Contract changelog
Every change to Tarno's REST API (https://api.tarno.cl) and MCP server (https://mcp.tarno.cl)
from 2026-09-17 onwards, newest first. REST and MCP are one contract: a route and its twin tool
(GET /trademarks ≡ search_trademarks) change together, and every entry names both.
The contract carries no version in the URL: this log is the version (see
Versioning & deprecation). The field-by-field schema lives in the live OpenAPI
reference, https://api.tarno.cl/docs.
How to read an entry
### date · [LABEL] summary, what changed and what you have to do. The labels:
| label | meaning | what you have to do |
|---|---|---|
[BREAK] |
breaking: a route, field, parameter or tool is removed, renamed or re-typed | change your code before it affects you; the entry says how |
[BEH] |
behaviour: same shape, different values or semantics | re-read the field's documentation; your code keeps compiling |
[ADD] |
addition: a new route, tool, field or optional parameter | nothing, beyond ignoring fields you do not know |
[FIX] |
fix of a fault, without changing the shape | nothing |
[DATA] |
the corpus itself moved (a recovery, a backfill, a corrected series) | re-run any measurement that depended on the old figures |
[WITHDRAWN] |
an entry announced and retracted before it shipped: the contract did not move | nothing |
Internal changes with no effect a client can see are not listed here.
November 2026
2026-11-15 · [BEH] Trademark record: procedureHistory[] orders same-day steps by the flow of INAPI's procedure
GET /trademarks/{applicationNumber} ≡ get_trademark.
- Within one day, entries now follow the place each section code takes in the flow of INAPI's procedure (e.g. the acceptance for processing
009before the publication payment439, the grant158before the certificate1230). That place was learnt from INAPI's e-file indices. - Codes without a known place come after those with one, and the remaining ties follow the recording order, as before.
- The order across different dates does not change, and neither do the entries or their fields.
- Measured on 1,345 days with several steps in 770 applications: the new order on its own reproduces INAPI's on 71.2 % of the days and contradicts it on 7.4 %; the rest are ties the recording order decides. INAPI does not publish each row's position, so some days may still come out in a different order.
What to do: nothing if you show the list as it comes. Do not reorder it by annotationId: you would lose this order. If you compared the order with INAPI's index, measure it again.
2026-11-15 · [DATA] Holders: 601 former holders with no tax id retired in 566 marks after their total transfer (A07T)
The same routes and tools as the previous holders and A07T entry. The shape does not change.
- Applied 2026-11-15 at 18:41 UTC: 601 holders with no tax id retired in 566 marks, each with
retiredBy: {kind: "section", section: "A07T"}andretiredOn= the date of the latestA07Tof its mark. They leaveholdersand appear among the former holders. - Measured afterwards: 0 pending in that population.
- Left out: 12 marks with a holder of a twin name, 77 with no re-read after the
A07T, and 13,373 with noA07T. The 77 are fixed on their next read. - Control case: mark 1292017.
- No request to INAPI.
What to do: repeat any measurement that depended on the current holders (counts per holder, a party's portfolio, co-holder relations).
2026-11-15 · [BEH] Trademark record: procedureHistory[] splits same-day steps by writ number and repeats identical ones
GET /trademarks/{applicationNumber} ≡ get_trademark.
- Two steps of the same day and code that INAPI tells apart only by the writ number it prints (
reference, such asC/2023/…) are now two entries. Before, one was shown. - A step INAPI lists several times identically on the same day, with no text or number to tell the copies apart, appears as many times as it is listed. Every copy carries the same
annotationId. - Part of the history was stored before this change. There, such steps show once until they are recovered or the application is read again.
- Madrid Protocol registrations have no «Expediente inicial (*)» row of the e-file index: INAPI's public record page does not publish it, and it is not invented.
What to do: do not use annotationId as a unique key within procedureHistory[] (use the position). If you compared the number of entries with INAPI's index, expect more entries on days with several writs of the same code.
2026-11-15 · [BEH] Holders: after a total transfer (A07T), the former holder with no tax id stops being a current holder
GET /trademarks/{applicationNumber} ≡ get_trademark and the reads by a party's validity (GET /parties/{id}/trademarks ≡ list_party_trademarks, GET /parties/{id}/outcomes ≡ get_party_outcomes, GET /parties/{id}/relations ≡ list_party_relations and the party filter of the trademark search). Same shape.
When a mark has a total transfer (A07T) and the mark's INAPI record, read with all its holders, no longer names a holder that has no tax id (not even by name), that holder stops being current: it leaves holders and appears among the former holders with retiredBy: {kind: "section", section: "A07T"} and retiredOn = the date of the latest A07T. In relations it becomes an assigned_to edge to the holder that stays, not a co_holder. Only when someone that read names stays as a current holder; without an A07T, with a partial transfer (A07P) or with matching names, nothing changes. It applies on the next read of each mark; the marks already re-read will get a separate entry with the date and the figures.
What to do: nothing if you read holders. If you counted co-holders of a transferred mark, stop counting the former holder with no tax id.
2026-11-15 · [BEH] Trademark record: procedureHistory[] reproduces INAPI's e-file index
GET /trademarks/{applicationNumber} ≡ get_trademark. Until now procedureHistory[] carried only the steps of the procedure table. INAPI's e-file index mixes them by date with the numbered annotations («Anotación de … N°: n», «[N°Anotación:n] Resolución …»), and now procedureHistory[] does too.
- The record page's numbered annotations (filings such as a transfer or a renewal, and the resolutions on them) come in, in date order among the steps.
- A filing INAPI lists both as a step and as a numbered annotation shows once, as the numbered annotation: its
annotationIdis that row's. They are paired one to one by date and code, so two filings of the same code on the same day are two entries. When the annotation has noreference, it takes the step's. - The ordering rule does not change: date and, within a day, the order the rows were recorded.
annotations[]does not change.
What to do: if you built the index by joining procedureHistory[] and annotations[], stop: rows would repeat. If you counted entries or stored a filing's annotationId, expect more entries and the annotationId of the numbered annotation.
2026-11-15 · [ADD] Trademark record: annotationNumber on each procedureHistory[] entry
GET /trademarks/{applicationNumber} ≡ get_trademark. Each procedureHistory[] entry carries annotationNumber (integer or null): INAPI's annotation number of that row, the one cited in a filing. It is null on a step of the procedure table and on an annotation stored without its number (part of the history lacks it).
What to do: nothing. Use it to show «N°» next to the row, or ignore it.
2026-11-15 · [BEH] Section legends: OPO16 gets INAPI's name, 449 and 678 are no longer cut short
The text of three section codes changes on all their rows: sectionName, procedureHistory[].inapiText, the Spanish label (sectionLabel, actLabel) and GET /sections ≡ list_sections, the same over REST and MCP. OPO16 goes from «Oposición - Pruebas oponente» (OPO12's name, attributed to it by mistake) to «Oposición - Solicita se cite a oir sentencia», as INAPI prints it, and no longer shares OPO12's label. 449 goes from «Resolución de abando» to «Resolución de abandono» and 678 from «… del CPC en oposició» to «… del CPC en oposición»: one of INAPI's publications cuts them short and the others print them whole. The family of the three does not change. The description of inapiText now says it is the reference spelling of the code, not of each row: INAPI may print the same code with a provenance prefix («(automático)») or a different abbreviation.
What to do: if you compared or stored these texts, use the new ones. Better still, compare by code/seccion, not by the text.
2026-11-13 · [BEH] Trademark filters: denomination takes up to 512 characters
The cap of denomination on GET /trademarks, POST /trademarks/search and search_trademarks (and on the statistics, the application-number list, the renewal calendar and portfolio.refine) goes from 256 to 512 characters. Some real denominations are longer than 256 characters (the longest, 304), and with 256 they could not be searched whole. q stays at 256. Above 512: 400 validation_error (too_big), the same over REST and MCP.
What to do: nothing. If you trimmed denomination to 256 characters, you no longer need to.
2026-11-13 · [BEH] Trademark filters: caps on the catalogue and number filters
On GET /trademarks, POST /trademarks/search and search_trademarks (and on the statistics, the application-number list, the renewal calendar and portfolio.refine, which share the same filter), the filters now have caps: niceClass up to 45 values; status up to 14; signType up to 14 values of up to 20 characters; denomination up to 256 characters, like q; registrationNumber and each number of the applicationNumber batch up to 17 characters. Above a cap: 400 validation_error (too_big), the same over REST and MCP. No value that can match a trademark is left out: the caps are the number of Nice classes and of codes in each catalogue, the longest code and the longest number an INAPI number's shape accepts.
What to do: nothing if you send real values. Do not repeat values beyond the cap, and trim denomination text to 256 characters.
2026-11-13 · [BEH] Vigilance horizon and registration signals: the text no longer names the initial import
GET /system/vigilance-horizon ≡ get_vigilance_horizon: rule keeps only the two real branches (estado_diario exempt, everything else kept 2 months); horizonDays stays at 60. GET /trademarks/registration-signals ≡ get_registration_signals: the S5 signal note says the opposition count comes from the annotations of INAPI's record pages or the Estado Diario. Text only; no figure or deadline changes.
What to do: if you matched on the text of rule or of the S5 note, stop; use horizonDays and the structured fields.
2026-11-12 · [BEH] Class statistics: coverage.annotationSources[] lists two sources
GET /nice-classes/{niceClass}/stats ≡ get_class_stats. The list of annotation sources in coverage no longer includes the initial import, which is not a source: its annotations come from INAPI's record pages or the Estado Diario. buscador and estado_diario remain, with their usual descriptions, and the opposition note (actsShare.opposition.coverageNote) says the opposition is counted from the annotations of INAPI's record pages and the Estado Diario. No figure changes.
What to do: if you had a branch for legacy_mongo, remove it: that value of source no longer arrives.
2026-11-11 · [BEH] Trademark record: a procedureHistory[] step stored twice shows once
GET /trademarks/{applicationNumber} ≡ get_trademark. INAPI returns some steps of the procedure table with their observation on one read and without it on another, and each version was kept apart: the same step appeared twice in procedureHistory[], with the same date and code.
- A step without observation that shares date and code with a step with observation now shows once: the
annotationIdis the one of the version with observation and, when that version has noreference, the other's is used. - Two steps with different observations are still two steps.
annotations[]does not change: withdedupe=falseboth versions are still published.- New steps are no longer stored twice.
What to do: nothing. If you counted the steps of procedureHistory[], there may be fewer.
2026-11-05 · [FIX] Descriptions of the reference and the MCP tools without internal vocabulary
The descriptions of the API reference and of the MCP tools have been reworded to say what each signal means (trigram similarity, edit distance, phonetic similarity, meaning similarity with meaning vectors computed by Tarno) instead of naming internal parts of the engine. The 400 for the retired offset parameter no longer cites internal codes: its detail is now "offset was withdrawn from the contract: paginate with limit + cursor". No field, value, type or behaviour changes.
What to do: nothing, unless you compare the text of detail.
2026-11-05 · [BEH] Renewal: renewed_expiry_pending also before the expiry, without renewal_window_open
Follow-up of the change of the same day. GET /trademarks/{applicationNumber} ≡ get_trademark, and the same renewalStatus in the collections, /insights, the portfolio report, the renewal calendar and calculate_fees. A mark whose renewal INAPI had already granted, with the published expiry not yet updated, came out as window_open and with the renewal_window_open milestone while that expiry had not passed.
- As soon as the renewal grant (dated from the expiry − 6 months) is recorded,
renewalStatusisrenewed_expiry_pendingbefore and after the expiry, instead ofwindow_open. - Those marks publish no
renewal_window_open(as they already published nograce_end): they do not matchmilestoneKind=renewal_window_opennor appear in the renewal calendar. calculate_fees(renewal) prices nothing in that status before the expiry either: emptylines,nulltotal andtotalOmittedReason: "renewed_expiry_pending".not_dueis unchanged. No new value.
What to do: treat renewed_expiry_pending as "renewed" also when the published expiry has not arrived yet: do not show an upcoming-expiry warning for those marks.
2026-11-05 · [BREAK] Renewal: new status renewed_expiry_pending when the renewal is already granted and the expiry is not yet updated
GET /trademarks/{applicationNumber} ≡ get_trademark, and the same renewalStatus in the collections, /insights (groupBy=renewalStatus), the portfolio report, the renewal calendar and calculate_fees. Some marks have a renewal INAPI already granted while the published expiry is still the previous one; until now they came out as grace or lapsed, with a grace_end, as if the renewal were pending or the term had passed.
renewalStatusgains a fifth value,renewed_expiry_pending: the published expiry has passed and the mark has a renewal grant (arenewals[]entry withkind: "grant") dated on or after that expiry − 6 months. It replacesgraceandlapsed.- Those marks publish no
grace_end: they are not matched bymilestoneKind=grace_endnor listed by the renewal calendar.renewal_window_open,not_dueandwindow_opendo not change. - The
renewalStatusfilter accepts the new value.calculate_fees(renewal) prices nothing in that status: emptylines,nulltotal andtotalOmittedReason: "renewed_expiry_pending". - Once the expiry is updated, the mark returns to the usual rule by itself.
What to do: if your client types renewalStatus or totalOmittedReason, add "renewed_expiry_pending". Read it as "renewed; expiry pending update", not as "did not renew".
2026-11-05 · [ADD] Trademark record: procedureHistory[], INAPI's procedure table with the reference of each step
GET /trademarks/{applicationNumber} ≡ get_trademark. New field, always present ([] when there are no steps), identical with or without dedupe and with any expand: the steps of the procedure table INAPI shows for the application on the application's public record page at INAPI, without the Estado Diario rows and without the per-fact grouping of annotations[].
- Each step:
annotationId(the sameidas inannotations[]),date,code,inapiText(INAPI's wording of that code, the same assectionName) andreference, the resolution or writ number INAPI prints on the step (2026/82718,C/2026/123550,A/2026/1447). referenceisnullwhen INAPI prints none or the step has not been read since it is kept: today it is on most historical steps, and it fills only when the mark is read again or, in part, by a backfill.- The record page's annotations (transfers, renewals, liens…) are told apart from the steps only by INAPI's annotation number, which much of the history lacks: one may appear as a step.
- Ordered by date and, within a day, by the order the steps were recorded: INAPI's position is not stored.
What to do: nothing; ignore it if you do not use it.
2026-11-05 · [BEH] renewal_window_open and calculate_fees.renewal.legalBasis cite Ley 19.880 art. 25 too, like grace_end
Follow-up of the 2026-11-04 grace_end change. GET /trademarks/{applicationNumber} ≡ get_trademark, the collections, the renewal calendar
(GET/POST /trademarks/renewal-calendar ≡ get_renewal_calendar) and POST /fee-schedule/calculate ≡ calculate_fees.
The grace_end milestone already cited four norms; now the three sources of the renewal term (renewal_window_open,
grace_end and the calculator's renewal.legalBasis) cite the same four, in the same order: Ley 19.039 art. 24 (term),
Decreto 82/2022 art. 27 (regulation), Código Civil art. 48 (computation) and {norm: "Ley 19.880", article: "25", role: "computation"}.
Same shape, one more list element: dates, rule, anchor, basis and renewalStatus do not change.
What to do: if you validate the legalBasis of renewal_window_open or of the calculator's renewal against three fixed entries, admit the fourth.
2026-11-05 · [BREAK] Trademark record: resolutions that do not grant a renewal now come out as kind: "outcome" in renewals[]
GET /trademarks/{applicationNumber} ≡ get_trademark. renewals[].kind had two values, request and grant, and every act that was not a grant came out as request: the resolutions that do not grant (abandonment, withdrawal, rejection, correction…) of a renewal request appeared as if they were new requests, so a trademark with a single filing could show four «requests».
kindgains a third value,outcome: INAPI resolves the request without granting it. Those acts (the Estado DiarioA4section) are no longerrequest.requestis still the filing and the request observed and pending.grantdoes not change. No entry leaves the list andlastRenewalOndoes not change.
What to do: if your client types kind as "request" | "grant", add "outcome". To count requests, count only request; an outcome is the result of a request, not a request.
2026-11-04 · [BEH] Madrid: a designation granted after a WIPO cancellation received during the procedure publishes caducado
GET /trademarks/{applicationNumber} ≡ get_trademark and every collection. Same shape. When WIPO notified the cancellation or the renunciation of an international registration (MAD01, DEATH) while the designation of Chile was still pending, and INAPI granted it afterwards with no later act reviving it (a renewal, a correction, a new birth notification, a cessation or an additional series of the same number), the designation publishes status: caducado, with statusInapi: "registrada" and statusStale: true, and the milestone designation_ceased anchored on that notification (anchor.kind: "wipo_death", rule: "wipo_death_received_before_grant", the same legal basis as the cessation). The status filter, groupBy=status, the facets, sort=status and liveOnly follow it. Today four designations. The text of the notices wipo_designation_ended and wipo_refused_recording no longer says the status is the one INAPI reports.
What to do: nothing in your code; if you store the status of Madrid marks, read them again: some move from registrada to caducado.
2026-11-04 · [BEH] Annotations: an M11 and the row for the same act published by the record page now form a single fact
GET /trademarks/{applicationNumber}?expand=annotations ≡ get_trademark, and the other annotation reads with dedupe. Same shape. An M11 annotation ("Miscellaneous decisions") carrying an act (e.g. 664B) and the row the record page publishes under that same section code, for the same mark and the same day, used to appear as two separate facts (versionsOfFact: 1, no duplicates[]). They now appear as one: the M11 is the main row and the other travels in duplicates[]; nothing is lost. With dedupe=false both rows remain, now sharing an eventId. Only exact pairs are grouped (one M11 and one twin per mark, day and act); ambiguous cases stay as they were.
What to do: if you group by eventId ?? id, nothing; if you counted acts per row, count again (the act count drops for grouped pairs).
2026-11-04 · [BREAK] The end of the renewal grace period moves to the next business day when it falls on a Saturday, Sunday or holiday
GET /trademarks/{applicationNumber} ≡ get_trademark, every trademark of the collections (GET /trademarks ≡
search_trademarks, batches, reports, portfolio) and the renewal calendar (GET/POST /trademarks/renewal-calendar
≡ get_renewal_calendar): the grace_end milestone is the expiry + 6 months and, when that day is a Saturday,
Sunday or holiday, the next business day (the same calendar as opposition_close). INAPI confirmed in writing
that a renewal filed on that business day is on time. renewalStatus stays grace up to and including that day
(so do the renewalStatus filter and groupBy=renewalStatus), and POST /fee-schedule/calculate ≡
calculate_fees prices that day as the 6th month, not as lapsed. The milestoneKind=grace_end filter matches the
new date and the watch's watchEndsOn follows it. The shape of the milestone changes:
rulegoes from"6_months_after_expiry"to"6_months_after_expiry_next_business_day";legalBasisgains a fourth norm:{norm: "Ley 19.880", article: "25", role: "computation"}(renewal_window_openand the fee calculator keep three);reservationis gone, and the codenon_business_last_day_extended_by_officeleaves thereservation.codevocabulary;basiscan beestimated, withestimateReasons: ["outside_holiday_calendar"], when the date falls after the declared coverage of the holiday calendar (today, up to 2030-12-31): the real day is that one or later. Renewal-calendar rows can beestimatedfor the same reason.
What to do: stop showing the grace_end reservation and accept its absence; if you compare rule, use the new
value; if you validate legalBasis, allow the fourth norm; treat grace_end and the calendar rows as derived
or estimated (show the reason, as for opposition_close); and drop
non_business_last_day_extended_by_office from your list of reservation codes.
2026-11-04 · [FIX] Encumbrances: a duplicate release that INAPI withdrew or rejected no longer lifts a charge
GET /trademarks/{applicationNumber}?expand=encumbrances ≡ get_trademark, and the reports carrying encumbrances[]. Same shape. Where INAPI's resolutions carry the annotation number, each filing is resolved by its own number: a release filed twice on one day whose duplicate was withdrawn or rejected counts once, and no longer marks as released (sometimes with releaseAttributionCertain and releaseStatusCertain set to true) a charge still in force. A charge rejected or abandoned by its number is no longer published as in force, and one registered by its number is no longer published as abandoned. If a release's registration resolution names a charge that is not in force on the mark, whatever that release closes, and every other charge it could have taken, carries releaseAttributionCertain: false and, unless every charge of the kind is released, releaseStatusCertain: false.
What to do: nothing in your code; re-read any encumbrances you store: some charges change state or release date.
2026-11-04 · [ADD] Annotation feed: maxActAgeDays, an act-age bound
POST /annotations/search ≡ search_annotations: new optional parameter maxActAgeDays (integer 0..3650). Only acts that were at most N days old when the corpus learned of them: their date is not before the UTC day of detection minus N. An act dated after its detection always passes; one without date never does. It requires detectedAfter and works in the corpus-wide feed and with applicationNumber or applicationNumbers. The order, the cursors and meta.detectedThrough do not change, and no row inside the bound is skipped or repeated. In the feed each page reads a bounded stretch: it can bring fewer than limit results, or none, with next not null. With dedupe, everything served is also within the bound against its detectedAt.
What to do: if you read the feed to alert and drop old acts on your side, pass maxActAgeDays with your window and keep following next until it is null even when a page comes back empty: you will no longer have to read a historical load of old acts just to drop it.
2026-11-03 · [BEH] Search and similarity: missing channels are declared, and an empty partial result is a 503
GET /trademarks / POST /trademarks/search ≡ search_trademarks (with q, denomination, a configured text, similarTo or similarToTrademark) ; the reports built on the similarity engine (get_clearance_report, get_conflict_landscape, get_opposition_brief, get_registration_signals, get_filing_context) do not carry the new field, only the 503 below. New optional field meta.degradedChannels: the search channels (same vocabulary as matchedBy) that should have fed the page and did not (they failed, did not answer in time, or were left out so that a cursor walk keeps page one's channels); absent when every channel ran. With some channel down and no results from the rest, the answer is now the retryable 503 service_unavailable, never an empty 200. A next page whose first-page channel is now down, or a page that could not be read whole, is also a 503. In the clearance report, the fuzzy group is omitted when its channel failed instead of being published empty.
What to do: treat the search 503 as retryable; optionally show a notice when meta.degradedChannels is present.
2026-11-03 · [BEH] MCP: invalid arguments are free and answered in Problem Details; reading a resource spends quota
An MCP tool call whose arguments do not match the schema no longer spends quota or burst limit (like a REST 400), and its error comes in the same body as every other error: status: 400, code: "validation_error" and per-field errors[], instead of the plain text «Input validation error: …». Reading a resource with a REST twin (resources/read of iris://catalogs/sections, nice-classes and holidays/{year}) now spends one unit of quota and burst, like its twin REST route (GET /sections, GET /nice-classes, GET /holidays); with the quota spent it answers 429 quota_exceeded. Listing resources is still free.
What to do: if you read the text of the MCP validation error, read the JSON (code, errors[]); count resource reads in your usage.
2026-11-02 · [BEH] Composite reports: at most 2 at a time; the excess may get a retryable 503
GET /reports/clearance ≡ get_clearance_report, GET /reports/portfolio ≡ get_portfolio_report, GET /reports/opposition-brief ≡ get_opposition_brief, and the two operations that run one of them inside (GET /reports/registration-signals ≡ get_registration_signals and GET /reports/filing-context ≡ get_filing_context with type=opposition) are served two at a time; the rest wait their turn for at most 10 s and, if it does not come, answer the already-declared 503 service_unavailable (at once when the queue is full). The aggregations of GET /insights/trademarks ≡ get_trademark_insights and the facets of the listing without text may be served from a copy up to 10 minutes old, renewed as soon as the corpus is updated (dataAsOf). Shapes and values do not change.
What to do: treat a 503 from these reports as retryable, with increasing backoff.
2026-11-02 · [BEH] MCP: only POST is accepted (other methods get 405)
The MCP server speaks JSON-RPC over POST only. Any other method (GET with Accept: text/event-stream, DELETE…) now answers 405 Method Not Allowed with Allow: POST and the standard error body (code: method_not_allowed); /health and /metrics are unchanged. A tools/call with a name that is not a registered tool always returns a JSON-RPC error.
What to do: nothing if you use a standard MCP client: it treats the 405 as "the server offers no stream" and carries on over POST.
October 2026
2026-10-29 · [BEH] renewalStatus: "grace" now says the registration is still in force
renewalStatus (on the trademark record, every mark of the collection, search, batch, similarity, window, a party's
marks and the reports ≡ get_trademark, search_trademarks…) keeps its shape, its values and its rule: a mark in its
grace period still says grace. What changes is what grace means: until today the documentation said it did not
state the registration's legal status; from today grace = in force, in the renewal grace period. The
registration lapses for want of renewal only if the grace period ends without a renewal request (Ley 19.039 art. 24, 2nd paragraph); a
renewal runs from the expiry with no gap (Decreto 82/2022, art. 27, 2nd paragraph); the lapse is not retroactive
(INAPI Directrices, chapter 23). INAPI cites registrations in their grace period as an art. 20 h obstacle just as it
cites those in force. The label of groupBy=renewalStatus (GET|POST /insights/trademarks ≡
get_trademark_insights) and of the portfolio's byRenewalStatus changes from "In grace period" to "In force, in
renewal grace period" (es: «Vigente, en plazo de gracia de renovación»; pt: «Em vigor, em período de graça de
renovação»). lapsed still does not say that no renewal was requested. See Milestones.
What to do: if you showed grace as "uncertain status" or "expired but renewable", show it as in force with the
renewal pending; if you compared the literal label, accept the new one.
2026-10-28 · [BEH] Trademark detail and report: new key notices[], notices checked in the expediente and in the WIPO record
GET /trademarks/{applicationNumber} ≡ get_trademark and the trademark report: a new optional key, notices[], with facts checked by hand in INAPI's expediente and in WIPO's International Register that the public sources do not carry. Two kinds: wipo_designation_ended, WIPO recorded in the International Register that Chile is no longer designated, and wipo_refused_recording, WIPO did not record INAPI's statement of grant of protection because Chile is not among the designated Contracting Parties (it carries no cause; its source is WIPO's refusal notice).
- Shape:
kind,date(the date of the recording),cause(renunciation,cancellation,dependency—cancellation because the basic mark ceased to be protected, Madrid Protocol art. 6.4—,non_renewal),label(the notice in es-CL, en and pt-BR),source—the primary source: WIPO's recording, withreference(RNN/2024/11),recordedOn,internationalRegistrationNumberandmadridMonitorUrl(the Madrid Monitor link)— andcorroboration—the backing: the document of INAPI's expediente, withdocumentIndex(its position in the index),labelanddate—. - A notice changes nothing else:
status,statusInapi,statusStaleandmilestones[]keep saying what INAPI says. - The key is omitted (never
[]nornull) on every mark without a notice: every other answer is unchanged.
What to do: if you show a trademark's detail, show its notices next to the status, with their source.
2026-10-28 · [DATA] Three Madrid designations will carry notices in notices[] (pending)
Pending. The marks 991718309, 991888075 and 991725022 will carry in notices[] the notice that WIPO recorded that Chile is no longer designated: 991718309 renunciation RNN/2024/11 (2024-02-15), 991888075 renunciation RNN/2026/19 (2026-04-28) and 991725022 cancellation by dependency CBN/2025/15 (2025-04-11). 991888075 will also carry wipo_refused_recording: WIPO did not record Chile's grant (GP-I/2135499301/BB, 2026-08-18). Their status stays registrada. A later entry will give the date.
What to do: nothing; if you keep a copy of those marks' detail, refresh it when we announce the entry.
2026-10-28 · [BEH] Parties: kindSource has a new value, name_pattern
GET /parties ≡ list_parties, GET /parties/{id} ≡ get_party and every party embedded in a trademark. Same shape; kindSource accepts a fourth value.
name_pattern: thekind(naturalorlegal) was inferred from the shape of a name that carries neither a RUT nor a company suffix (words of personal names, or company words such asSOCIEDAD,FUNDACION,GROUP). It is an inference calibrated against the parties whose kind their RUT fixes, and it is only applied where the lower Wilson bound of its accuracy is ≥ 98 %.- A RUT that arrives later replaces it with
tax_id. - Pending: no party carries it today. When the classification is applied, tens of thousands of parties that are
kind: unknowntoday will becomenaturalorlegal. A separate entry will give the date and the figures.
What to do: if you validate kindSource against a closed list, add name_pattern. If you need certainty about the kind, require kindSource: tax_id.
2026-10-28 · [BEH] Parties: a party's own RUT no longer enters its name
GET /parties ≡ list_parties, GET /parties/{id} ≡ get_party and every party embedded in a trademark. Same shape; name was already nullable.
- INAPI sometimes prints a RUT in the applicant's name box. From today, the party's own RUT (the one it carries as
taxId) does not entername: a name that was only that RUT isnull, and a name with more text keeps the text.publishedNameis still INAPI's text, RUT included. - A row whose name is only a RUT binds to the party of the trademark that carries that RUT; if it is a holder row and the trademark already has a company as current holder, it creates no new party (it was the company with its representative's RUT in the name box).
- The name lists of a trademark summary (
holders,representatives,formerHolders,formerRepresentatives) leave out a party without a name instead of returning an empty string. - Pending: about 70 parties whose name is a RUT today will change
name(companies take the legal name of their RUT), 2 will be merged into the party that carries their RUT, and 39 false holder links of 31 "RUT" parties will be removed (those trademarks keep the company as holder). A separate entry will give the date and the figures.
What to do: if you displayed name, display publishedName when name is null.
2026-10-28 · [BEH] Estado Diario, type-5 bulletin: a new D5A annotation carries the date of the act, not of the bulletin
Section D5A (acts on registered marks): annotations[] of the trademark, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin. Same shape; what changes is the meaning of date for this section.
- The rule. The
dateof aD5Aannotation is the date of the act, the one INAPI's record page gives it, not the date of the bulletin that printed it. INAPI sometimes prints in the type-5 bulletin the acts of 1 to 4 days earlier: 31 bulletins between 2025-05-22 and 2026-09-22, and it still happens (measured 2026-10-28). - From today, when the record page already holds the act's annotation number for that mark, dated before the bulletin, the new row carries that date. Otherwise it carries the bulletin's date, as before. An act printed in the bulletin of day B may appear dated B−1 … B−4. An act already stored is never written twice.
get_daily_bulletinfor day B lists theD5Aacts dated B; thealtascoverage row keeps counting what that bulletin published, so on those days the two can differ. Expected, not a gap.
What to do: if you group or count D5A acts by date, keep in mind the date is the act's.
2026-10-28 · [DATA] Estado Diario, type-5 bulletin: 2,281 D5A annotations will move to the date of their act (pending)
Section D5A: the same routes and tools as the entry above. Same shape.
- Pending. 2,281 existing
D5Aannotations of those 31 bulletins will move to the date of their act, 1 to 4 days earlier (onlydatechanges). Only those the record page dates without doubt; the others stay as they are until the record page dates them. No movement and no watch notice. A separate entry will give the date and the real figures.
What to do: recount any D5A acts by date when we announce the entry.
2026-10-28 · [BREAK] Madrid: INAPI's cessation after a WIPO DEATH notification publishes caducado and the milestone designation_ceased
GET /trademarks/{applicationNumber} ≡ get_trademark and every mark of the collections (GET /trademarks ≡
search_trademarks, batches, reports): a Madrid designation that INAPI declared ceased after a WIPO DEATH
notification —cessation 1238 (renunciation of Chile), 1240 (cancellation), 1241 (non-renewal) or 1242
(dependency on the basic mark), after the latest DEATH and the latest PROLONG notification, with no later
registration entry— now publishes status: caducado, with INAPI's literal status in statusInapi (today,
denegada) and statusStale: true. 1237 (partial transfer: the protection continues in the new series) does
not count. The status= filter, groupBy=status, the facets, sort=status and liveOnly follow the same
rule. The mark also carries a new milestone, designation_ceased: basis: "derived", date = anchor.date,
anchor {kind: "office_cessation", date, annotationId, sectionCode} (the cessation act),
rule: "office_cessation_after_wipo_death" and legalBasis Madrid Protocol arts. 6.3, 6.4 and 9 bis and Common
Regulations rules 25 and 27, with the new role extinction. Not filterable. Breaking only for a client with closed
lists: milestones[].kind and legalBasis[].role grow. It affects about 160 Madrid marks today. See
Milestones.
What to do: accept designation_ceased and extinction in your types; if you counted those marks as
denegada, they now come as caducado (statusInapi keeps INAPI's value).
2026-10-28 · [BREAK] The late-renewal surcharge is computed with the Directrices 2026 table: the total is no longer null in the grace period
POST /fee-schedule/calculate ≡ calculate_fees: for a renewal paid within the grace period
(renewal.renewalStatus: "grace"), the month row of INAPI's Directrices 2026 surcharge table (chapter 13; chapter 23
§42: 1st month 6 UTM, up to 12 UTM in the 6th, per class) applies as the rule in force and the total is computed:
totalUtm and totalClp carry the figure and totalOmittedReason is null (before: null with "surcharge_doubt").
Example: 3 classes paid one day after expiry ⇒ totalUtm: 18; in the 2nd month, 21.6; in the 6th, 36. New field
renewal.surchargeRuleSource: "Directrices 2026"; renewal.ruleIsOpenQuestion becomes false; totalOmittedReason
can now only be "lapsed". The discrepancy with the «(*)» note and with other INAPI fee publications is still published,
as an informative note that blocks nothing: surchargeDoubt (in GET /fee-schedule ≡ get_fee_schedule and in
renewal) keeps its name and its literal notes[], and its description now states the rule. subjectToDoubt and
hasDoubt keep their values. See official fees and the surcharge note.
What to do: show the total in the grace period too, citing the source (surchargeRuleSource) and the note; if you
handled totalOmittedReason: "surcharge_doubt", that case no longer arrives.
2026-10-28 · [BEH] Natural persons' RUT: the service can publish it masked (today, whole)
Natural persons' RUT is still published whole and can still be searched: nothing changes today. What is new is that the service can switch to another mode, announced in advance, in which:
- the
taxIdof a party whosekindis notlegal(natural, andunknownby prudence) is masked:***plus the last 3 digits of the body and the check digit (12345678-5→***678-5).*never appears in a real RUT. Legal entities do not change. It affects the parties ofGET /parties≡list_parties,GET /parties/{id}≡get_party,GET /trademarks/{applicationNumber}≡get_trademark, relations, the portfolio, the opposition briefs and the filing context; - the
taxIdfilter (parties, trademarks and annotations) finds legal entities only: a natural person's RUT returns an empty result, not an error; - every collection's
metacarriesnaturalPersonTaxId: "masked"(andGET /trademarks/application-numbers≡list_application_numbers, which has nometa, carries it at its top level). As long as that key is absent, nothing changes.
What to do: nothing today. To be ready: treat a taxId containing * as masked (show it as is, do not
validate its check digit or use it as a key or filter); identify a party by its id, never by its taxId.
2026-10-28 · [DATA] Estado Diario, type-5 bulletin: 210 D5A acts of 2024-11-13 and 2025-01-06, and eight days without an own bulletin (origin: not_published)
Section D5A (acts on registered marks): annotations[] of the trademark, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin; and GET /system/coverage ≡ get_system_coverage. The shape does not change.
- From today, the reader reads the bulletins that print each act's registration number on a line of its own, above the row (the 211 rows «without a registration column»), and checks the mark by INAPI's annotation number (
annotationNumber): if the number and the registration name different marks, the row is not written. Nothing changes for a bulletin printed the usual way. - Pending. 210
D5Aacts will be added (100 dated 2024-11-13 and 110 dated 2025-01-06), each on its mark, checked by the annotation number and by the registration (measured 2026-10-28: 210 of 210 agree). Those two days will gain theiraltascoverage row. One row of 2026-09-24, with no registration printed, stays out. - Pending. Eight days on which INAPI did not publish an own type-5 bulletin (that day's PDF repeats an earlier one, or does not exist) will get their
altasrow withacts: 0and the neworiginvaluenot_published: 2024-05-29, 2025-05-16, 2025-06-16, 2025-08-13, 2025-10-01, 2025-10-09, 2025-10-23 and 2025-10-30. No annotation is added: those days' acts are already there, through the record page. - The new rows will carry the day they are added as their detection date (
detectedAfter). No movement and no watch notice will be generated. A separate entry will give the date and the actual figures.
What to do: if you read the coverage origin, accept not_published (a day without an own bulletin, not an ingestion gap). When we announce the lot, repeat any count of D5A acts or altas days.
2026-10-28 · [BEH] The end of the renewal grace period carries its reservation: grace_end with reservation
GET /trademarks/{applicationNumber} ≡ get_trademark and every mark in the collections (GET /trademarks ≡
search_trademarks, batches, reports): the milestone grace_end (expiry + 6 months, Ley 19.039 art. 24) now
carries reservation, with the new code non_business_last_day_extended_by_office and its text in Spanish,
English and Portuguese. The date does not change: it is the expiry plus exactly six months and is not moved
when it falls on a Saturday, Sunday or public holiday. The reservation says that since October 2024 INAPI has
accepted renewals filed on the next business day (Ley 19.880, art. 25, third paragraph), so the published date is
the last safe day, not the last possible one. Same shape: reservation already existed on grant_fee_close;
reservation.code gains one value. Also in the renewal calendar (GET/POST /trademarks/renewal-calendar ≡
get_renewal_calendar): every grace_end row carries the reservation (renewal_window_open rows do not). See
Milestones.
What to do: show the grace_end date with its reservation, as you do for grant_fee_close; if you
validate reservation.code against a closed list, add the new value.
2026-10-28 · [FIX] Documentation: the anchor of opposition_close is the last publication of the extract
Documentation only: no value changes. When INAPI sets the publication of an extract aside and republishes it,
it counts the 30 business days of the opposition term from the last publication, and that is the date it
publishes as the mark's publication date. opposition_close ≡ oppositionWindow.deadline are already anchored
on it (anchor.date); the milestones table now says so.
What to do: nothing; if you computed the close yourself from the first publication, use the milestone's
anchor.date.
2026-10-26 · [BREAK] The date the application was decided: the milestone resolution in milestones[]
GET /trademarks/{applicationNumber} ≡ get_trademark and every trademark of the collections (GET /trademarks ≡
search_trademarks, batches, searches) carry a new milestone kind, resolution: the date of the act by which
INAPI accepted or rejected the application (not a deadline). basis: "derived", date = anchor.date, anchor
{kind: "acceptance" | "rejection", date, annotationId, sectionCode}, rule: "latest_acceptance_or_rejection"
and legalBasis Ley 19.039 art. 13 with the new role notification (the resolution is notified by the Estado
Diario). The most recent acceptance or rejection act counts, including the acts that comply with a TDPI ruling
granting the registration. Absent when the application was never decided (e.g. 428 «tener por no
presentada»), when a later TDPI act alters it without saying which way, or on a renewal file. Not filterable.
Breaking only for a client with closed lists: milestones[].kind and legalBasis[].role grow. See
Milestones.
What to do: accept resolution and notification in your types; if you dated an application's decision
yourself, read the milestone.
2026-10-26 · [BEH] MCP: explain_section_code explains three more codes
MCP only. explain_section_code now explains 33 codes: 428 (resolution deeming an application not filed because its extract was not published in time; it ends the procedure without accepting or rejecting it), 961 (grant of a pre-approved renewal, used from 2014 to 2017, with no prior acceptance act) and W231 (acceptance of a renewal in the pre-2012 nomenclature) are added. Before, they answered known: false / no_se_sabe. The explanations of the other 30 codes do not change; knowledgeRevision moves to the new revision in every answer. Same shape.
What to do: nothing; if you cached a no_se_sabe for any of these three codes, ask again.
2026-10-26 · [ADD] The renewal calendar of a portfolio: POST /trademarks/renewal-calendar with portfolio
New POST /trademarks/renewal-calendar ≡ get_renewal_calendar: the same answer as the GET by the JSON body,
plus the portfolio object (party, refine, includeApplicationNumber, excludeApplicationNumber), the same
one as POST /trademarks/search and POST /trademarks/application-numbers. A mark added by hand is in the
calendar even when it fails refine; one excluded by hand never is. The cursor is bound to the whole filter,
portfolio included (with another, 400 validation_error on cursor). With portfolio, a top-level party,
includeApplicationNumber or excludeApplicationNumber is a 400. The GET does not change.
What to do: nothing; for the calendar of a portfolio with marks added by hand, use the POST (or
get_renewal_calendar) with portfolio (see Reports).
2026-10-26 · [BEH] MCP: explain_section_code explains four more codes
MCP only. explain_section_code now explains 30 codes: C12C (rulings in non-use cancellation proceedings), C13L (final rulings in cancellation and territorial-limitation proceedings; the name is ours, not INAPI's literal heading, and the explanation says so), C16C (miscellaneous resolutions in non-use cancellation proceedings) and MSO08 (assignment of a pending application, which does not say to whom) are added. Before, they answered known: false / no_se_sabe. The explanations of the other 26 codes do not change; knowledgeRevision moves to the 2026-10-26 revision in every answer. Same shape.
What to do: nothing; if you cached a no_se_sabe for any of these four codes, ask again.
2026-10-25 · [ADD] What the SII says about a party: sii on GET /parties/{id} and GET /parties
GET /parties/{id} ≡ get_party and every party of GET /parties ≡ list_parties carry a new field,
sii: for a party with a Chilean RUT listed in the SII roster of legal entities, the legal name, the company
type (SII code and label in es-CL/en/pt-BR), the start of activities, the end of business, the comuna (with its CUT code) and
region of the current domicile per the SII (the street address is not published) and the current branches by
comuna. Each block carries source: "sii" and asOf = the SII file date. The SII domicile is the current
one, not the one at filing (territory). sii is null without a Chilean RUT, when the SII does not list
the party, or when the party was merged into another. Nothing that already existed changes.
What to do: nothing; read sii if it helps (see People search).
2026-10-24 · [BEH] Holders of a mark whose pending application was assigned (MSO08), and co_holder relations
When an application still pending is assigned (INAPI code MSO08, which does not say to whom), the
holder who assigned it stops being a current holder: in GET /trademarks/{id} ≡ get_trademark it
leaves holders and appears among the former holders with retiredBy: {kind: "section", section:
"MSO08"} and retiredOn = the date of the assignment; if INAPI's open data names it as representative,
it moves to representatives and is not listed as a former holder. Only when the data is unambiguous: a
truncated or differently written name is left alone. In GET /parties/{id}/relations ≡
list_party_relations, a holder retired this way is not a former co-holder (co_holder) of the one who
stayed. Not a transfer: assigned_to/assigned_from and groupBy=transferYear do not change. Same shape. That holder counts as
former in every read by validity: GET /parties/{id}/trademarks ≡ list_party_trademarks and
GET /parties/{id}/outcomes ≡ get_party_outcomes with relation=former, and the party filter of the
trademark search finds it only with includeFormer=true.
What to do: nothing; if you store holders, re-read the marks with an assigned application.
2026-10-23 · [BEH] A party's review.reason: new code holder_tax_id_on_representative
review.reason (on GET /parties/{id} ≡ get_party and on embedded parties) gains a value,
holder_tax_id_on_representative: a representative reached INAPI carrying the RUT of the mark's own
titular. That RUT is no longer attached to the representative, which is identified by its name, and
the titular stops also appearing as a representative (ambas) of its own mark.
What to do: if you compare review.reason against a closed list, add this code.
2026-10-21 · [ADD] POST /trademarks/search ≡ search_trademarks, GET /trademarks and /trademarks/application-numbers ≡ list_application_numbers: the text search configuration object
New optional field text, with Algolia's vocabulary: { query, queryType, typoTolerance,
minWordSizefor1Typo, minWordSizefor2Typos, exactOnSingleWordQuery, removeWordsIfNoResults, phonetic,
semantic, getRankingInfo }. With every option at its default it answers exactly what q answers.
Configured, the answer has two blocks: first the keyword block, a deterministic and complete set
(queryType prefixNone/prefixLast/prefixAll/contains/exact, typoTolerance
true/false/min/strict with the minimum word sizes, removeWordsIfNoResults
none/lastWords/allOptional), ordered exact → prefix → more words → fewer typos → sort →
number; then, with phonetic or semantic, the approximate hits (in relevance order). Without
approximate hits total is exact; with them total is absent, meta.keywordTotal is the keyword
block's exact size and meta.estimatedTotal counts both (a lower bound). At most 16 words. getRankingInfo: true adds rankingInfo to
each hit. In a GET it is flattened: text=…&textQueryType=…&textTypoTolerance=…. In
/trademarks/application-numbers and in portfolio.refine.text it works as a set: it requires
phonetic: false and semantic: false. text with q, denomination, the batch, similarity or the
opposition window is a 400, and so is minScore with a configured text. New code
422 text_set_too_large: more than 50,000 candidates in the keyword block.
What to do: for a literal search (a portfolio's search box), send
text: { query, queryType: "contains", typoTolerance: false, phonetic: false, semantic: false }; for
the exact → prefix → approximate order, configure text instead of q.
2026-10-22 · [ADD] Annotations and encumbrances: INAPI's annotation number (annotationNumber)
Every annotation (get_trademark with annotations, get_annotation, search_annotations, the report)
carries annotationNumber: INAPI's «N° de anotación», the one cited in a filing. A resolution carries the
number of the annotation it resolves. Every encumbrances[] entry carries annotationNumber (the charge)
and releaseAnnotationNumber (the paired release). null when unknown: the Estado Diario does not print
it, and older history is being numbered progressively.
What to do: nothing; use it to cite the annotation. Do not invent one where it is null.
2026-10-22 · [BEH] Encumbrances: two identical charges of one day are two entries, and a release is paired with the charge INAPI says it lifts
encumbrances[] (GET /trademarks/{applicationNumber}?expand=encumbrances ≡ get_trademark, and the
report) now uses INAPI's annotation number. Two charges of the same kind, day and text (two pledges
to the same creditor filed together) were one entry and are now two; charges of one day to different
creditors too. When INAPI states which charge a release lifts, the pairing is no longer a convention:
releaseAttributionCertain and releaseStatusCertain come out true. Marks whose history does not
carry numbers yet read as before, and are updated progressively.
What to do: nothing in your code (same shape). A mark may show one more active charge than
yesterday: a charge that existed and was not visible.
2026-10-21 · [ADD] Encumbrances: releaseStatusCertain says whether a charge's state is certain
Every entry of encumbrances[] (GET /trademarks/{applicationNumber}?expand=encumbrances ≡ get_trademark)
gains releaseStatusCertain (boolean). It answers "is this charge still in force?": true when its
state (active or released) is certain, even where releaseAttributionCertain is false because
which release lifted which charge is only an attribution. releaseAttributionCertain keeps its meaning.
Two rules make the state certain under an uncertain pairing: when every charge of the kind on the
mark is released, each one is released for certain (only its releasedOn stays an attribution); and a
charge filed on the same day as a release is never the one that release lifted. false whenever
releaseObservable is false. About four in five released charges that carried
releaseAttributionCertain: false now carry releaseStatusCertain: true.
Known limit: when INAPI recorded two identical charges of one kind on the same day and the record shows only one, the state can be wrong; the same is true of charges certain before this change.
What to do: to say whether a charge is in force, read releaseStatusCertain; keep
releaseAttributionCertain for whether releasedOn is that charge's own release date.
2026-10-21 · [FIX] GET /trademarks/renewal-calendar ≡ get_renewal_calendar: the cursor is only valid with the filter it was issued for
A cursor obtained with one filter and used with another answered 200 with a page that was the next of
nothing. The cursor is now bound to the filter (every field but limit and cursor; from and to
included, limit may change): with another filter, 400 validation_error with one entry on cursor, the same
in REST and MCP and the same pattern as /trademarks/application-numbers. A cursor issued before this change
is unreadable (same 400).
What to do: request every page with the same filter and the previous next; on the 400, restart without
cursor.
2026-10-21 · [BEH] A party's review.reason: new code representative_tax_id_on_holder
review.reason (in GET /parties/{id} ≡ get_party and in embedded parties) gains one value,
representative_tax_id_on_holder: the holder reached INAPI carrying the RUT of a law firm or lawyer that
represents it. That RUT is no longer assigned to the holder, which is identified by its name, and the firm
no longer shows as co-holder of its client's mark.
What to do: if you compare review.reason against a closed list, add this code.
2026-10-19 · [ADD] POST /trademarks/search and POST /trademarks/application-numbers ≡ search_trademarks and list_application_numbers: the portfolio object
An application added by hand to a portfolio no longer drops out for failing the portfolio's refinements.
New optional field portfolio, { party, refine, includeApplicationNumber, excludeApplicationNumber }:
result = (((parties ∩ refine) ∪ included) − excluded) ∩ the top-level filters. refine takes the
predicate filters of the list of numbers and narrows only the parties' trademarks; the top-level filters
are the listing filters and apply to everything, included applications too. Inside portfolio each list
takes up to 10,000 numbers. With portfolio, a top-level party, includeApplicationNumber or
excludeApplicationNumber is a 400 validation_error; without it, nothing changes. New route
POST /trademarks/application-numbers: the same answer as the GET (cursor, limit, 409
cursor_stale), with the filter in the body; the cursor is bound to the whole filter, portfolio
included. GET /trademarks stays flat.
What to do: for a portfolio, send its parties, its refinements and its lists inside portfolio, and
keep only the listing filters at the top level.
2026-10-19 · [BREAK] GET /trademarks/application-numbers ≡ list_application_numbers: page by page, with a cursor; truncated is removed
The list of numbers of a filter is no longer cut at 10,000: it is walked page by page. New parameters
limit (1..10,000, default 10,000) and cursor; the answer carries next, the cursor of the next page
(null on the last one), and no longer carries truncated (next not null replaces it). total
(exact, for the whole set), cap and dataAsOf stay, and total and dataAsOf are the same on every
page of one walk. The order does not change (number as text) and no number repeats or goes missing
between pages. The cursor is only valid with the same filter (another one is a 400 validation_error on
cursor) and with the corpus of page one: if a sync started or finished between pages, the next one
answers 409 cursor_stale.
What to do: stop reading truncated; while next is not null, ask for the next page with
cursor=<next> and the same filter. On a 409 cursor_stale, start again without cursor.
2026-10-19 · [BEH] GET /parties ≡ list_parties no longer returns merged parties; GET /parties/{id} ≡ get_party adds mergedInto
When a party record is merged into another one (they were the same party), its trademarks and annotations
move to the one that survives. The party search no longer returns the merged record, under any sort or
filter (it used to come out with trademarkCount: 0). GET /parties/{id} on it still returns it, with a
new field, mergedInto: the id of the party that holds its trademarks now (the end of the chain if there
were several merges). For any other party, mergedInto is null. The other reads by id (/trademarks,
/relations, /outcomes, the portfolio report) answer for the record asked for, which has no trademarks
after the merge.
What to do: if you store party ids, when you read one whose mergedInto is not null, replace it
with that value.
2026-10-18 · [ADD] GET /trademarks/application-numbers ≡ list_application_numbers: every application number of a filter, at once
A new route and its twin tool: the application numbers of every trademark that matches the filter of
GET /trademarks (batch, party, includeApplicationNumber/excludeApplicationNumber, niceClass,
status, parties, dates, milestones, renewalStatus…), ordered by number as text (not numeric), in a single answer with no
pagination: { applicationNumbers, total, truncated, cap, dataAsOf, coverageThrough }. At most 10,000
numbers (cap); when more match, the list holds the first 10,000, truncated is true and total is
the exact count. q, denomination, similarTo, similarToTrademark, allClasses,
oppositionWindow, facets and the norm filters are a 400 validation_error naming the field; there is
no limit, cursor, sort or minScore. Scope brands:read.
What to do: nothing; use it when you need the whole set behind a filter and only the numbers. If
truncated is true, narrow the filter.
2026-10-18 · [FIX] %, _ and \ in a substring filter match literally
holder and representative (GET /trademarks ≡ search_trademarks and every other operation that
uses the trademark filter), denomination as a substring (insights and the renewal calendar),
GET /parties?q= ≡ list_parties and GET /legal-norms?q= ≡ list_legal_norms: a value containing
%, _ or \ returned nothing; now each of those characters matches itself and only itself
(holder=A_B finds «A_B 50% LTDA», not «AXB…»). Values without them do not change. Same shape.
What to do: nothing.
2026-10-18 · [BEH] Monthly quota and watch caps: those of your organisation's plan
Every organisation has a plan. Its monthly quota may apply to all its keys together, on top of
each key's own quota: whichever runs out first applies, with the same 429 quota_exceeded and its
Retry-After, and the RateLimit-* headers and quota in GET /system/me describe that one. The daily
cap of watch additions (429 watch_daily_limit_reached) also becomes the plan's, as the cap of
applications watched at once already was. Today every organisation is on the usual plan: the numbers do
not change (no organisation quota, 500 watched, 500 additions a day).
What to do: nothing; if your integration reads RateLimit-Limit, keep reading it: it is the limit that applies to you.
2026-10-17 · [BEH] GET /watch/hits ≡ list_similarity_hits: a watched mark of 1–2 characters only gets short matches
If the watched denomination has 1 or 2 letters or digits («Q», «TE», «+B»; signs and spaces do not count),
the watch only stores matches of 1 to 3 letters or digits, at any score: the identical ones and the
short ones that sound alike («Q» → «K», «W» → «V»). It leaves out the long names that carry that letter or
an initial that sounds like it («TE» → «THE JUNGLE BOOK», «G» → «G T D INTERNET»), which used to come in at
score 1 and gave hundreds of matches per acronym. Watched marks of 3 characters or more do not change.
Same shape, fewer rows. The similarity search (similarTo, similarToTrademark), the opposition window and
availability do not change.
What to do: nothing; if you watch short acronyms, expect far fewer matches.
2026-10-16 · [BEH] GET /trademarks ≡ search_trademarks: a comma no longer separates names in holder, representative and taxId
A value of holder, representative or taxId is one whole name, with its commas and other signs:
holder=PEPSICO, INC. searches for that name, not for PEPSICO or INC. as until now. For several
names (OR), repeat the parameter (?holder=A&holder=B) or use one array item per name in the body of
POST /trademarks/search and in MCP. Also, in those three filters: leading/trailing spaces are trimmed and
runs of spaces count as one; an empty or whitespace-only value, one with fewer than 2 characters or one
without a letter or digit is a 400 validation_error (an empty value used to drop the filter and
holder=% ended in a 504). taxId still requires the shape of a RUT (400 invalid_rut). Also in
/insights/trademarks and the renewal calendar. The lists of numbers, ids and vocabulary (niceClass,
status, party, includeApplicationNumber…) still accept a,b. And GET /legal-norms?q= ≡
list_legal_norms treats % and _ as characters, not wildcards.
What to do: if you sent several comma-separated names in one holder, representative or taxId,
repeat the parameter instead. Encode each value (the comma travels as %2C, a + as %2B).
2026-10-16 · [ADD] GET /trademarks ≡ search_trademarks: include and exclude applications by hand in a portfolio
New parameters includeApplicationNumber and excludeApplicationNumber, repeatable or comma-separated
(up to 500 each): result = ((party filters ∪ included) − excluded) ∩ every other filter. The party
filters (party, holder, representative, taxId) form the portfolio scope; the included
applications are added to it and the excluded ones are always removed, even when they match. Every other
filter (status, class, text, dates…) applies afterwards, to the included ones too. Without a party
filter, the included applications are the portfolio. Order, cursor, total and facets, on the final
set. Also on POST /trademarks/search, /insights/trademarks and the renewal calendar. A malformed
number, an empty list or more than 500 is 400 validation_error; with oppositionWindow, a 400.
What to do: nothing; a portfolio with applications added or removed by hand is now one request.
2026-10-16 · [ADD] GET /trademarks ≡ search_trademarks: party filter by party id and role
New parameter party=<id>:<role>, repeatable (up to 20): the trademarks of some parties by their
internal id (the one of GET /parties?q=), each with its role, holder or representative.
Several values are OR'ed; party is AND'ed with every other filter, and relation chooses current
(default), former or all relations. A role means what it means in GET /parties/{id}/trademarks?role=,
so party=<id>:representative returns the same trademarks as that route, now filterable, sortable and
with facets. It works for parties without a RUT. Also on POST /trademarks/search,
/insights/trademarks (≡ get_trademark_insights) and the renewal calendar (≡ get_renewal_calendar).
A malformed value or more than 20 is 400 validation_error; with oppositionWindow, a 400.
What to do: nothing; for a portfolio of parties chosen by id, use party instead of the name.
2026-10-16 · [BEH] GET /watch/hits ≡ list_similarity_hits: between 0.7 and 0.8 the watch only stores matches that carry a word of the watched mark
The watch still stores only matches with score ≥ 0.7, and from 0.8 up nothing changes. Between 0.7 and
0.8 it stores a match only if it carries a word of the watched denomination, equal or one letter away
(«NOMADES» → «NOMADE», «CMP» → «CMPC»). The words that count are the watched mark's own words that are rare
in its classes (fewer than 50 marks), without its generic terms; if it has none, any of its own counts.
Digits glued to letters are split before comparing («NOMADE98» → «nomade 98»). It leaves out matches that
only share an ordinary word of the trade («SL Abogados» → «Abogados Cordillera»): 9 in 10 of that band.
Same shape, fewer rows; matches already stored follow the rule from this change on. The similarity search
(similarTo, similarToTrademark), the opposition window and availability do not change.
What to do: nothing; if you count matches per watched application, expect fewer between 0.7 and 0.8.
2026-10-16 · [ADD] GET /trademarks ≡ search_trademarks: every trademark carries publishedOn and registeredOn
Every item of the trademark collection (listing, search, batch by number, similarity, opposition window and
the trademarks of a party, and the trademarks inside the reports) now carries publishedOn (publication
date) and registeredOn (registration date), YYYY-MM-DD or null, with the same value as the detail
GET /trademarks/{applicationNumber} ≡ get_trademark. They are the dates you could already sort and filter
by. Other fields are unchanged.
What to do: nothing; if you sort by those dates, you can show and check the order without fetching each trademark's detail.
2026-10-16 · [ADD] GET /trademarks/{applicationNumber}?expand=encumbrances ≡ get_trademark: which date recordedOn is, and which releases were attributed by age
Every encumbrances[] entry carries two new fields; no existing field changes value.
recordedOnBasis("recordal","filing"ornull) says which daterecordedOnis. On a modern charge (A0xI) in force or released it is"filing": the date is the annotation's filing, not its registration. On a 2007–2011 charge it is"recordal": its registration date.nullwhenrecordedOnisnull.releaseAttributionCertainisfalsewhen a release arrived while two or more charges of its kind were in force and was paired with the oldest one: the data does not say which it lifted. Every charge in force at that moment carries it: the one listedreleasedmay still be in force, and anactiveone may be the one lifted.
What to do: do not print "recorded on…" over a recordedOn with recordedOnBasis: "filing" ("filed
on…"). With releaseAttributionCertain: false, do not state which creditor is still in force: say that
releases were recorded and that the data does not say which charges they lifted.
2026-10-15 · [BEH] GET /trademarks/{applicationNumber} ≡ get_trademark: a Madrid base number with no Chilean expediente of its own answers its lettered expediente
Some international registrations have, in Chile, only a lettered expediente (<base>-A): the base
registration never designated Chile. Those trademarks are published under the lettered number
(applicationNumber = <base>-<letter>, with baseApplicationNumber and seriesSuffix), with all their
data and history. Asking for the base number in GET /trademarks/{applicationNumber} ≡ get_trademark,
in /deadlines ≡ get_deadlines and in /annotations ≡ list_trademark_annotations answers 200 with
that trademark instead of 404. The batch lookup by number, search and watch do not recognise the base
number.
What to do: if you store the base number of one of those trademarks, use the applicationNumber the
detail returns (the lettered number).
2026-10-14 · [BEH] POST /annotations/search ≡ search_annotations: detectedAfter and detectedUntil require seconds
Both filters follow RFC 3339, which mandates seconds when the time carries Z or an offset.
2026-10-01T10:00Z (no seconds) used to be accepted and is now a 400 validation_error;
2026-10-01T10:00:00Z and 2026-10-01T10:00:00.000Z keep working. Every instant the contract returns
(meta.detectedThrough, dataAsOf, …) already carries seconds, so the cursor pattern is unchanged.
The OpenAPI document also now publishes niceClass of GET /reports/clearance, /reports/conflict-landscape
and /reports/registration-signals as required (required: true): it already was, omitting it gave 400.
What to do: if you send detectedAfter/detectedUntil without seconds, add :00. If you generate a
client from the OpenAPI document, regenerate it: niceClass becomes required on those three routes.
2026-10-14 · [FIX] stale discounts only the Sunday hours that have elapsed
GET /system/freshness ≡ get_system_freshness, only without thresholdHours. Until now the whole Sunday
was discounted from 00:00: a source that was already stale on Saturday (Saturday's synchronization did not
arrive) looked fresh again at Sunday 00:00 and stayed so until the afternoon. Now only the Sunday time that has
already passed is discounted: on Sunday the age stops growing and on Monday it resumes where it stopped.
- A normal weekend reads the same: a Saturday ingestion is not
stalethrough Sunday nor on Monday until 24 h net later. - A source
staleon Saturday staysstaleall Sunday. - Same shape;
stalenessSecondsis still the real age; with an explicitthresholdHoursnothing changes.
What to do: nothing.
2026-10-14 · [BEH] Trademark record: each renewal says whether it is the request or the grant, and lastRenewalOn counts only grants
GET /trademarks/{applicationNumber} ≡ get_trademark. renewals[] counted the renewal request (the filing) as a renewal, so a trademark whose renewal was requested and then abandoned or rejected showed that date as its «last renewal».
- Every
renewals[]entry carries a new field,kind:request(the renewal was requested: filed, or observed or in progress; not yet granted on that date) orgrant(INAPI grants, accepts or records it). No entry leaves the list. lastRenewalOnis now the date of the latestgrant, omitted when there is none. Measured on 2026-10-14 over 212,708 trademarks withlastRenewalOn: 208,921 do not change, 1,265 move back to an earlier grant and 2,522 lose it (1,891 of them lapsed).
What to do: if you showed lastRenewalOn as «last renewal», you can now do so. If you render renewals[], tell kind apart: a request is not a renewal done.
2026-10-11 · [ADD] The published OpenAPI says what its version means and carries info.x-revision
GET /docs/openapi.json. info.version is still the date of the last cut of the contract, not a change counter: it does not move with every entry of this log. The document now says so in info.description and carries a new field, info.x-revision: the date of the newest entry of this log with a visible effect on the document ([INT], [DATA] and [WITHDRAWN] do not count). The ETag of the response, which already existed, identifies its exact content. The MCP server does not change: its version is still the same cut date.
What to do: to know whether you must re-read the contract, compare info.x-revision (or, better, the ETag with If-None-Match) with what you stored, not info.version. If x-revision is missing, do not assume it.
2026-10-11 · [BEH] Text search: it stops waiting for a channel that has not answered in 3 s
GET /trademarks ≡ search_trademarks and POST /trademarks/search with q or denomination (without similarTo). The shape does not change.
A text search queries four channels at once (exact, fuzzy, phonetic and by meaning) and until now waited for all four: one slow channel could stretch the answer beyond 10 s. From today, once 3 s have passed and at least one channel has already found matches, the search answers 200 with the channels it has. The channel left out behaves like one that failed: its matches are not in the page, its label is not in matchedBy or matchedFields[].kind, and if it is the meaning channel signals.semantic may be null. It never returns an empty page or an error because of this.
What to do: nothing. If you repeat the same search a few seconds later it may bring more matches and a slightly different order; do not mix pages from two runs separated by a long wait. The latency objective of the search is still p95 < 3 s (SLA).
2026-10-11 · [DATA] Estado Diario: 9,492 annotations from 2024-09-30 to 2026-09-02 recovered
Annotations of a trademark (GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin). The shape does not change.
Applied on 2026-10-11 between 01:02 and 01:23 UTC. They are publications of INAPI's daily bulletin the corpus did not have:
- 2,657 of nullity, caducity and cancellation proceedings (sections
C9NtoC16N, and four ofC16C,C12CandC13L), which were not read until September 2026. - 2,420 of trademarks with a six-digit application number (filed between 1995 and 2012), nearly all «Resoluciones varias» (
M11). - 4,097 second resolutions of the same day for the same trademark and section (nearly all
M11): only one per day was kept before. - 318 of other
A,CandMsections. - They are born with detection date 2026-10-11: if you page by
detectedAfteryou will see them as newly detected, with adatebetween 2024-09-30 and 2026-09-02. Filter bydateif you only want what is recent. - The four annotations of
C16C,C12CandC13Lcome withsectionName: nulluntil those codes are named. - No new trademarks, no movements, no watch alerts.
What to do: repeat any measurement that depended on the annotation count of that period; if you consume the detection feed, filter by date.
2026-10-11 · [BEH] Sections: a new family, cancelacion_y_limitacion_territorial, and three Estado Diario codes stop going nameless
GET /sections ≡ list_sections, and sectionFamily/sectionName/sectionLabel on every annotation. The shape does not change (family and sectionFamily stay text or null), but a value that did not exist before appears: cancelacion_y_limitacion_territorial. Families go from nine to ten.
| code | name | family | annotations |
|---|---|---|---|
C16C |
Resoluciones varias en juicios de Caducidad | caducidad |
2 |
C12C |
Fallos de juicios de Caducidad | caducidad |
1 |
C13L |
Ejecutorias en juicios de Cancelación y de Limitación Territorial | cancelacion_y_limitacion_territorial |
1 |
- Those 4 annotations (measured in production on 2026-10-11 at 01:46 UTC, read-only) stop carrying null
sectionNameandsectionLabel, andsectionFamilygoes fromnullto the family. family=caducidadgoes from six to eight codes;family=cancelacion_y_limitacion_territorialreturnsC13L.- ⚠ INAPI prints every
C*Lsection twice, once for Cancelación and once for Limitación Territorial, and the code does not say which. The single row came out under Cancelación; the name and the family name both proceedings on purpose. It is our label, not the literal heading. - No code changes family:
family=oppositionand the others return what they did.meta.dataAsOfofGET /sectionsmoves to 2026-10-11. - Left without a family:
CAN08andCAN99(cancellation of a GI/AO, another matter) and the administrative cancellations of a registration (CANTstays inresolucion).
What to do: if you validate family or sectionFamily against a closed list, add cancelacion_y_limitacion_territorial and treat any value you do not know as another family, not as an error. The current list is read from GET /sections.
2026-10-10 · [DATA] Estado Diario: 354 annotations of 16–21 May 2024 recovered
Acts on registered marks, section D5A (GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin). Same shape.
INAPI did not publish this bulletin on 16, 17 or 20 May 2024: it published a single one, dated the 21st (a holiday), with the acts of those days. Applied on 2026-10-10 at 23:09 UTC: 354 annotations come in, each under the date of its act — 82 on 2024-05-16, 139 on 2024-05-17 and 133 on 2024-05-20.
- 37 carry the business day before the date INAPI gives them: 12 acts of Saturday 18 and Sunday 19 go with Friday 17, and 25 of Tuesday 21 (a holiday) with Monday 20. This is what INAPI does in its ordinary bulletins.
- They carry a detection date of 2026-10-10: if you page by
detectedAfter, you will see them as newly detected, with adateof May 2024. - No new marks and no watch alerts.
What to do: nothing, unless a measurement of yours depended on the annotation count of those days.
2026-10-10 · [DATA] Estado Diario: 22 annotations of 2026-03-30 recovered
Annotations of a trademark (GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin). Same shape.
Applied on 2026-10-10 at 21:10 UTC. 22 annotations dated 2026-03-30 appear: 12 of Madrid-series marks whose number the bulletin printed split over two lines, and 10 from nullity sections that were not read at the time. No new marks and no watch alerts.
- They carry a detection date of 2026-10-10: if you page by
detectedAfter, you will see them as newly detected.
What to do: nothing, unless a measurement of yours depended on that day's annotation count.
2026-10-10 · [DATA] Estado Diario ledger: 553 days from 2024-04-19 to 2026-08-06 gain their altas bulletin row (applied 2026-10-10 at 21:21 UTC)
GET /system/coverage ≡ get_system_coverage and bulletins[] of GET /reports/daily-bulletin ≡ get_daily_bulletin. The shape does not change.
The ledger of processed days had no altas bulletin row for the days before 2026-08-07, although their D5A acts are already in the corpus. 553 days (61,376 acts) were seeded, from 2024-04-19 to 2026-08-06, with acts = the number of D5A acts of the day and origin = seed. After the seeding, 596 days with D5A have their altas row (65,945 acts, 2024-04-19 → 2026-10-08).
- Only 2025-12-13 (73 acts, a repeated bulletin with no type-2 row) still has none.
- No trademark or annotation data changes. A row with
origin: "seed"is rebuilt from the corpus: it does not mean «read by the daily sweep».
What to do: nothing, except re-running any count of altas days or acts from the coverage ledger.
2026-10-10 · [BEH] Sections: new family caducidad for the acts of cancellation-for-non-use proceedings
GET /sections ≡ list_sections, and sectionFamily on every annotation. The acts of a cancellation-for-non-use proceeding ("juicio de caducidad") now form a family of their own. The shape does not change (family and sectionFamily are still text or null), but a value that did not exist before now appears: caducidad. The families go from eight to nine.
- Six codes go from
nulltocaducidad:C11C"Citaciones a oír sentencia en juicios de Caducidad",CAD04"Caducidad - Certificación de notificación por Receptor u otro ministro de fe",CAD06"Caducidad - Contestación del traslado",CAD08"Caducidad - Demanda de caducidad del registro",CAD10"Caducidad - Otras peticiones" andCAD99"Caducidad - Varios (sin nro de demanda)". That is 21 annotations of 16 trademarks, from 2022-07-19 to 2026-02-23. - The
family=caducidadfilter returns them, andGET /sections?family=caducidadlists the six codes. - No code changes family:
family=oppositionand the others return what they returned before. Class statistics, reports, deadlines and statuses do not change. - Still without a family: the counterclaims of caducidad filed inside an opposition or a nullity action (
1121,1122,OPO24,OPO26,NUL30,NUL31). caducidadis not the trademark statuscaducado: the family names acts of a proceeding, and the trademark under attack may still be registered.
What to do: if you validate family or sectionFamily against a closed list, add caducidad and treat any value you do not know as another family, not as an error. The current list is read from GET /sections.
2026-10-10 · [BEH] Watch: a deactivated organisation is no longer watched for, and gets its list back if it is reactivated
GET /watch/applications ≡ list_watched_applications, GET /watch/hits ≡ list_similarity_hits. The shape does not change, and an active organisation notices nothing.
- While an organisation is deactivated (its keys answer
401), the service stops watching for it. Its list is kept. - If it is reactivated, the list comes back by itself within a minute, with the
watchedSinceit had and without spending the daily cap. The applications only it watched go throughbaselineStatus: "pending"for a while. - In
GET /watch/hits, what it can read again comes after the cursor it kept, with the sameids: it will receive again matches it had already read, plus the new ones.
What to do: nothing if your organisation is not deactivated. If it is reactivated, wait a minute before reading the list and handle repeated matches by id.
2026-10-10 · [BREAK] Two new deadlines with their article: answering the transfer of an opposition and paying the grant fees
GET /trademarks… ≡ search_trademarks / get_trademark (milestones[]), GET /reports/filing-context ≡ get_filing_context and GET /trademarks/{applicationNumber}/deadlines ≡ get_deadlines. Closed lists grow, get_deadlines now tells its entries apart by kind, and one name changes in get_filing_context.
opposition_response_close(milestone, deadline ofget_deadlinesand term ofget_filing_contextwithtype=opposition_response): the last day to answer the transfer of an opposition, the 30th business day after it was notified, not extendable (Ley 19.039 art. 9 and art. 11; Ley 19.880 art. 25; Decreto 82/2022 art. 24). Not answering is not abandonment.derivedwhen the corpus has the Estado Diario notification and the transfer resolution;estimated(notification_inferred_from_resolution_dateortransfer_resolution_not_in_corpus) when one of the two is missing.grant_fee_close(milestone, and deadline ofget_deadlines): the last day to prove payment of the grant fees, the 60th business day after the notification of the acceptance or of the certificate of finality; unpaid, the application is held abandoned (Ley 19.039 art. 18 bis E and art. 11; Ley 19.880 art. 25). Alwaysestimated(anchor_by_office_practice) and with a new field,reservation {code, text {es, en, pt}}: the law counts from the day the resolution becomes final and INAPI from its notification; the date follows what INAPI applies.- The
anchorof these two milestones also carries the act the date comes from (annotationId,sectionCode). - In
get_deadlines,deadlines[]andomitted[]now carry threekindvalues (appeal_close,grant_fee_closeandopposition_response_close), each with its own fields: theappeal_closeentries do not change, but they are no longer the only ones. The answer deadline carriesestimateReasons([]whenderived) andanswerObserved(the firstOPO05from the anchor on, a fact). A mark whose latest act is a partial acceptance now has the payment inomitted[]withacceptance_not_final, in addition to its appeal deadline. - In
get_filing_context, the omitted termopposition_response_term(no_legal_basis_in_knowledge_base) is no longer emitted: it is calledopposition_response_closeand, when not given, the reason isno_transfer_on_record,madrid_designation,prior_legal_regime,anchor_on_non_business_dayoroutside_holiday_calendar. There is a new fact,opposition.transfer. - Both deadlines exist only for national marks anchored from 2022-05-09. Neither filters with
milestoneKind. The answers to office actions on a filing (661B,813B,931,925) still have no date: they have no article.
What you have to do: in get_deadlines, filter deadlines[] and omitted[] by kind before reading their fields (do not read position 0); in get_filing_context, replace opposition_response_term with opposition_response_close. If you validate milestones[].kind, estimateReasons, deadlines[].kind or omitted[].kind against a closed list, widen it. Show grant_fee_close with its reservation, never as a certain date, and an estimated one as an estimate.
2026-10-10 · [BEH] Party relations: the opposition ones (opposed, opposed_by) cover from 2024-04-19
GET /parties/{id}/relations ≡ list_party_relations. Since the corpus reads the Estado Diario bulletins of 2024-04-19 to 2024-09-27, the oppositions of that period are linked to their parties: of the 12,402 annotations with parties in those months, all 12,402 have their opponent linked (in 4,998 the opponent is identified as a party; 40.3 %, against 43.2 % from 2024-09-30 on).
coverageFromofopposedandopposed_bybecomes2024-04-19(was2024-09-30). The shape does not change; the other five types stay as they were. The opposition blocks of the reports and ofget_filing_contextdeclare the same date.- A party that opposed, or was opposed, between 2024-04-19 and 2024-09-27 gains those edges:
firstOn,lastOn,trademarkCountandevidencecan change. An opponent whose printed name matches no party or several still makes no edge (meta.unresolvedOpponentMentions). - The
actsShare.opposition.coverageNoteofGET /nice-classes/{niceClass}/stats≡get_class_statsnow says "since 2024-04-19"; the statistics do not change.
What to do: "never opposed" now reads "not since 2024-04-19". If you stored opposition edges or your own coverage date, read them again.
2026-10-10 · [BEH] Trademark report: renewals[] stops repeating the renewal of an establishment that the Estado Diario already publishes
renewals[] of the trademark report. The renewal of an establishment mark that also extends its territorial coverage came out twice: once as section A5 of the Estado Diario (act AREPTLT or ARENTLT) and once as resolution 1108 (commercial establishment) or 1123 (industrial, as class-40 services) of the record page. It is the same act, on the same mark and day: of the 1,196 A5, 989 have their 1108 or 1123 that same day (all 176 before 2024-09-30).
1108and1123count as the same act as anAREPTLTof the other source, same mark and day: about 959 fewer entries across the corpus.lastRenewalOndoes not change (same date).- The 28
A5withARENTLTand a1108or1123still come out twice: the pair is not guessed.annotations[]andeventIddo not change.
What to do: if you counted a mark's renewals from renewals[], count them again; the renewal of these establishments is no longer counted twice.
2026-10-10 · [DATA] Estado Diario, annotations of the type-5 bulletin: bulletins that print the annotation number next to the act are now read too
Every annotation view (annotations[] of the trademark report, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin). The shape does not change.
Nine bulletins of 2026 (17 and 18 February; 8, 10, 21, 29 and 30 April; 4 May; 9 July) print the annotation number and the act in a single text, and their annotations on registered marks (renewals, transfers, changes of name…) were not read: those nine days have none.
- From today, a new bulletin printed that way is read like any other. Nothing changes for a bulletin printed the usual way.
- The nine past days are now recovered (2026-10-10, 21:06 UTC): 997 new annotations, dated between 17 February and 9 July 2026. In
detectedAfterthey show up as detected on 10 October: they are not new acts, and they raise no movements and no alerts. - On 13 August 2025 INAPI did not publish this bulletin: that day's annotations exist only in each mark's history.
What to do: if you treat detectedAfter as «recent act», also filter by the date of the act.
2026-10-10 · [DATA] Estado Diario: a Madrid-series application number printed over two lines (7 + 2 digits) is read whole; 22 annotations of 2026-03-30 pending recovery
Annotations of a trademark (GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin). Same shape.
INAPI's daily bulletin sometimes prints the number of a Madrid-series mark (nine digits) split over two lines. When the split left seven digits above and two below, the publication was not read. Measured over the 616 archived bulletins (2024-04-19 → 2026-10-09): it happened in one bulletin, that of 2026-03-30, with 12 publications. No wrong data was published; those 12 annotations are missing.
- From today such a publication is read under its full number.
- Pending: recovering 22 annotations of 2026-03-30 (the 12, plus 10 from nullity sections that were not read at the time). A separate entry will announce it. No new marks and no watch alerts.
What to do: nothing.
2026-10-10 · [DATA] Estado Diario, new applications of the type-5 bulletin: the trademark name is read from its own column even when INAPI changes the column widths
denomination of a newly filed trademark (GET /trademarks/{n} ≡ get_trademark, GET /trademarks ≡ search_trademarks and every view that shows the name). The shape does not change.
A new application first arrives through INAPI's daily bulletin and is completed, hours later, with its record. The bulletin does not always print its columns with the same width: measured over 618 bulletins (2024-04-19 → 2026-10-08), in 201 some column was shifted and in 9 the table was not read. On those days, until the record arrived, the new trademark could show its Nice classes as its name, an empty name, or not show at all.
- From today the new trademark shows its name from the first moment, on those days too. Nothing changes for a bulletin printed the usual way.
- What is already published is right except for one trademark. INAPI's record replaces the name when it arrives, and it had arrived for all but one: a three-dimensional trademark (no text) that shows its class as
denomination. It was corrected on 2026-10-10 at 19:25 UTC (emptydenomination; its text vector was deleted and the next embeddings delta creates the new one).
What to do: nothing.
2026-10-10 · [DATA] Estado Diario, section D5A: 111 acts of a repeated bulletin move from 2025-10-09 to 2025-10-10 (applied 2026-10-10 at 19:53 UTC)
Corrects, for one pair of dates, this same day's entry «945 acts that a repeated bulletin left under two dates were retired». That clean-up kept the first date in the eight repeated bulletins. A later measurement shows that INAPI records the act under the first date in all but one: the 111 acts of the pair 2025-10-09 / 2025-10-10 carry the date 2025-10-10 in INAPI's record page (111 of 111). The bulletin INAPI serves under the date of the 9th carries the table of the 10th.
- Applied on 2026-10-10 at 19:53 UTC (lot
d5a-boletin-repetido-par-261010): 111 rows restored with theirid, 111 twins retired. After it, 2025-10-09 has 0 bulletinD5Arows and 2025-10-10 has 111. - The 111 annotations retired on 2026-10-10 at 10:22 UTC under the date 2025-10-10 exist again, exactly as they were: same
id, same fields. - Their 111 twins dated 2025-10-09 are retired: their
idstops existing andGET /annotations/{id}answers404. 2025-10-09 is left with noD5Aannotation. - The total number of annotations does not change (+111 −111). No trademark changes, and this generates no watch notice.
- About 80 annotations that the record page dates 2025-10-09 are in no bulletin: they stay as record-page annotations of their marks, without a
D5Arow.
What to do: if you store D5A annotations of 2025-10-09 or 2025-10-10, or their id, read them again: it is applied.
2026-10-10 · [DATA] Legacy-import marks: 32,500 read from INAPI's record page; they gain category, sign type and registration and expiry dates
Until today, the marks that arrived on 2026-06-29 only with the old system's import (filed between 1995 and 2008, plus 812 from Madrid) had never been read from the record page. The load finished at 06:24 UTC on 2026-10-10: 32,500 marks read, 32,500 requests (one per mark), all with result cambio, no foreign record, none recovered by search and no retry pending. No denomination or registration number changed with the visit. It affects GET /trademarks/{applicationNumber} ≡ get_trademark, the renewal calendar, GET /nice-classes/{niceClass}/stats ≡ get_class_stats and the detectedAfter feed.
- Every mark read goes through the usual writer: it gains its category, sign type and registration and expiry dates when the record page reports them, and its
updatedAtmoves. 1 mark remains without a category (it was 27,301 after the fill of the 2026-10-09 entry). - Annotations later than the old system's last read appear in the detection feed (
detectedAfter) as newly detected, with their real date, which may be months old. - The 135 Madrid base numbers with a lettered expediente were frozen before the load and were not read.
- The shape does not change.
What to do: if you keep a copy of these marks, read them again; repeat any measurement of class statistics or expiry dates that counted them as without category or date.
2026-10-10 · [BEH] Class statistics: the substantive objection counts the decisions of before 2012 that cite a ground; W152 and W147 join family resolucion
GET /nice-classes/{niceClass}/stats ≡ get_class_stats, and the section catalogue (GET /sections ≡ list_sections, sectionFamily in the annotations). The shape does not change.
actsShare.substantiveObjection. Until February 2012 INAPI did not publish the substantive objection as an act of its own: from before, only the final decision remains. From today the figure counts, besides the acts of family fondo, the trademark whose final decision of that period — the refusal W160 «Rechaza Resolucion (Plazo 15 Apelacion)» or the partial acceptance W152 «Aceptada parcial (c/limitacion)» — cites article 19 or 20 of Law 19.039 (the grounds for refusal). A W160 that cites neither, for example one that refuses a renewal, does not count.
| filing year | before | now | trademarks |
|---|---|---|---|
| 2009 | 160 (0.3 %) | 3,599 (6.8 %) | 52,975 |
| 2010 | 590 (1.0 %) | 4,341 (7.5 %) | 57,501 |
| 2011 | 6,670 (9.8 %) | 6,940 (10.2 %) | 67,859 |
| 2004–2008 | 12 | 137 | 6,208 |
| 2012 onwards | unchanged | unchanged |
(All classes; a trademark counts once per class.) Per class and for 2009–2011, for example: class 35, from 834 to 1,621 of 13,376; class 41, from 544 to 1,274 of 10,515.
- The figure of those years is a lower bound: an objection the applicant overcame left no act. The note
substantive_objection_history_limited, the field'scoverageNoteanddefinitionsay so with the new text, in the three languages. The code of the note does not change; it is now also emitted when the period starts in 2011 (before, only when it started before 2011): the trademarks of 2011 gain decisions too and their figure is also a lower bound. - The default period (the last five years) does not change, so neither does the substantive-objection signal of
GET /reports/registration-signals≡get_registration_signals. dataAsOfof the statistics moves to the moment of this change.
Section catalogue. W152 «Aceptada parcial (c/limitacion)» (499 annotations) and W147 «Fallo Notif. Aceptada Parcial (15 ds.Ap)» (310): sectionFamily goes from null to resolucion, the family of their current equivalents 444 and 485. The filter family=resolucion returns 809 more annotations. W160 stays resolucion, and family fondo does not change: it is the examination (formulating the objection and resolving it); the final decision is resolucion.
No deadline, alert or status changes.
What to do: repeat any measurement of actsShare.substantiveObjection that includes years before 2012, and do not present it as comparable with later years. If you store section families, read GET /sections again.
2026-10-10 · [DATA] Estado Diario, second table of the type-5 bulletin: three row layouts the reader was skipping are now read; 1,007 acts of past days were added on 2026-10-10
Section D5A (acts on registered marks): annotations[] of the record, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin. Same shape.
Measured over the 617 type-5 bulletins of the archive (2024-04-19 → 2026-10-07): 1,321 rows in 92 bulletins were still not read. From today three layouts are read: the registration number printed with six digits, without its leading zero (637 rows); the act wrapped over two lines (232); and the page set with other margins (241).
- From today, a bulletin printed in any of those three layouts is read like the others. Nothing changes for a bulletin printed the usual way.
- Applied on 2026-10-10 at 10:21 UTC (lot
boletines-archivo-261010-formas): 1,007 newD5Aannotations over 78 days were added, from 2025-09-11 to 2026-09-28: 473 with the six-digit registration, 232 with the act on two lines, 231 from pages with other margins and 71 more rows of the Saturday 2025-12-13 bulletin (73 that day in all), which had been left out of the 2026-10-09 recovery. None was completed, and the 7,100 rows of those days that were already there were not touched. Annotations: 11,476,209 → 11,477,216 (+1,007). Marks (846,141), the record page queue andmovement_logdid not move. - Not read: 211 rows of 3 bulletins (2024-11-13, 2025-01-06 and one of 2026-09-24) that do not print the registration number: without it there is no safe way to tell which mark the act belongs to.
- The new rows carry the day they were added (2026-10-10) as their detection date (
detectedAfter), with adatein 2025 and 2026. No movement and no watch notice was generated.
What to do: filter by date as well or ignore that batch in the detection feed, and repeat any measurement that counted D5A annotations of those days.
2026-10-10 · [DATA] Estado Diario, section D5A: 945 acts that a repeated bulletin left under two dates were retired: only the first stays
Twelve type-5 bulletins carry, under their own date, the second table of an earlier bulletin: the same acts, with the same INAPI annotation number. Eight of them were ingested, so the same act appears twice, with two dates. The first date is the good one.
- Applied on 2026-10-10 at 10:22 UTC (lot
d5a-boletin-repetido-261010): the copy on the later date was retired, 945 annotations, of 2025-05-16 (113), 2025-05-20 (81), 2025-06-16 (110), 2025-09-10 (125), 2025-10-01 (171), 2025-10-10 (111), 2026-06-05 (93) and 2026-07-03 (141). Each had its exact twin (same mark, same section, same text) on the first date (2025-05-15, 05-19, 06-13, 09-09, 09-29 or 09-30, 10-09, 2026-06-04 and 07-02), which stays; none had anything hanging from it. Annotations: 11,477,216 → 11,476,271 (−945). - The
idof a retired annotation stops existing:GET /annotations/{id}answers404. - The other four repeated bulletins (2025-10-23, 2025-10-30, 2025-11-20 and 2026-09-14) were never ingested and stay out: their acts are already there under the first date.
- No trademark changes, and this generates no watch notice.
What to do: if you store D5A annotations of those eight dates, or their id, read them again.
2026-10-10 · [BEH] No synchronization on Sundays: dataAsOf does not advance that day and stale does not count Sundays
GET /system/freshness ≡ get_system_freshness, meta.dataAsOf and the «sync finished» notice. Same shape;
the cadence changes. The daily synchronization stops running on Sundays (Chile time; INAPI does not publish
the Estado Diario on Sundays). From Monday to Saturday it runs as before.
- On Sunday
dataAsOfdoes not advance (it stays on Saturday) and nosync.finishednotice is sent. On Monday the synchronization picks up everything accumulated and nothing is lost. staledoes not raise a false alarm because of Sunday: withoutthresholdHours, every Chile Sunday between a source's last ingestion and now is taken off its age before it is compared with the threshold (24 h). If Monday's does not arrive the source turnsstale; an outage during the week is still flagged after the usual 24 h.stalenessSecondsis always the real age. With an explicitthresholdHoursthe threshold means exactly that, with no allowance.
What to do: nothing, except not to expect a new dataAsOf or a sync.finished webhook on Sundays. If you
set thresholdHours, add the Sunday yourself (e.g. 54).
2026-10-09 · [DATA] Madrid registrations with divisionals: the record of the base number no longer receives another series' data
GET /trademarks/{applicationNumber} ≡ get_trademark. For up to 58 Madrid Protocol registrations, the
source answered a query for the base number with the data of a divisional application of the same
registration, and the record of the base number showed them. From today an answer that does not belong
to the number asked is not stored: the record shows its own data when the source delivers them, and
otherwise keeps what it had. No shape change.
What to do: if you keep a copy of Madrid Protocol marks (9-digit numbers), read them again; classes, status, registration number and history may change for those few marks.
2026-10-09 · [DATA] Estado Diario from 2024-10-03 to 2026-10-07: 10,159 acts of the second table of the type-5 bulletin that the daily sweep was not reading are added
The daily sweep did not read the second table of the type-5 bulletin (acts on registered marks, section D5A) when the registration number fell outside its column. Measured on our own archive, 511 type-5 bulletins from 2024-09-30 to 2026-10-07: 111 with the table out of band, 10,775 rows lost, the first on 2024-10-03. In production 10,159 new annotations are added over 106 days (2024-10-03 → 2026-10-07; annotations[] of the record, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin).
- Annotations: 11,463,897 → 11,474,056 (+10,159).
D5Aannotations dated from 2024-10-01: 43,057 → 53,216. No duplicates. No trademark is created and no existing annotation is completed. - Of the 10,775 rows, 60 repeat the same mark and act within the lot, and 556 are not added: four bulletins that repeat the acts of the previous business day (2025-10-23, 2025-10-30, 2025-11-20 and 2026-09-14; 484 rows) and the Saturday 2025-12-13 bulletin (72 rows). Separately, 1,321 rows that no reader reads yet. They are decided apart.
- From the 2026-10-08 bulletin on the daily sweep reads that table by itself (that day: 130 acts, in band).
- The daily counters of additions for 15 of those days, which already had a row, now add up to 1,252 (
/system/coverage).
⚠ The new rows were created on 2026-10-09, the date detectedAfter uses, so 10,159 old acts show as detected today although their date is from 2024 to 2026. No movement, no record-page request and no watch alert was generated.
Same shapes, no field added or removed.
What to do: also filter by date or ignore the 2026-10-09 batch. Repeat any measurement that counted D5A annotations since 2024-10-03.
2026-10-09 · [BEH] Watch: the series alert also arrives when you start watching and in the daily review
The direct alert of a series of the registration you watch (relation: "series" in GET /watch/hits ≡ list_similarity_hits) no longer depends only on the pass that follows a load of series. Same shape, same fields and same values; what changes is when it arrives. No additional series is loaded yet: no response changes today.
- When you start watching an application of a Madrid registration — the base number or a series —, normally once its first search finishes you already have the other live series of the registration, whether or not they share a class with yours; in every case, at the latest with the next daily review (for example, if you watch again an application you had recently stopped watching, its search is not repeated). Before, only the ones sharing a class arrived at once.
- A series that was in a terminal status and revives is announced in the next daily review. Before, it could be left without an alert.
- The daily review also announces any other live series of your registration that was not in your list yet. It can take up to a day.
originkeeps its three values and still says who wrote the item: a series alert can carrybaseline(the first search),sync(the daily review) orseries_load(the pass that follows a load). It does not tell an alert from a match:relationdoes.- Unchanged: one item per watched application and series, never the series you already watch nor the base number, never a series in a terminal status, "your own portfolio is not a threat" and the cursor. An alert you already had is not repeated and keeps its
id.
Details: Direct alert: a new series of the registration you watch.
What to do: nothing if you tell the alert apart by relation === "series". If you detected it with origin === "series_load", switch to relation.
2026-10-09 · [BREAK] Watch: direct alert of a new series of the registration you watch — relation: "series", score can be null, origin gains series_load
Whoever watches an application of a Madrid registration — its base number (991652209) or one of its series (991652209-2) — gets in GET /watch/hits ≡ list_similarity_hits one item per other additional series of that registration (991652209-3), even when it shares no class with the watched one: the similarity search, which works inside the classes of the watched application, would never have found it. No additional series is loaded yet: no response changes today. It is tagged [BREAK] because one field is re-typed and one vocabulary grows.
relation(new, optional; the only value is"series"): present only when the application inmatchedis an additional series of the same registration as the watched application (its base number or another series). On an ordinary match the key is absent.score: fromnumbertonumber | null.nullonly on an item withrelationthat the similarity engine did not find; no score is invented, andsignalsis then{}. Withoutrelationit is always a number.origin: gainsseries_load(written by the pass that follows a load of series), next tobaselineandsync.- If the series also shares a class and the engine finds it, it is the same item (one
id), withrelation: "series"and its realscore: never two items for the same series. - The cursor, the
applicationNumberfilter, "your own portfolio is not a threat" and closing on a terminal status work the same. It is not closed for "no longer resembles": it only loses itsscore. - Whoever watches the base number and whoever watches another series of the registration get the alert. It is one item per watched application: if you watch both
991652209and991652209-2, series991652209-3reaches you twice, once under each (like any match of two of your watched applications). The series you already watch is never announced, and neither is the base number: only additional series are news. - A match without
relation, and the whole response of anyone who watches no registration with series, do not change.
Details and example: Direct alert: a new series of the registration you watch.
What to do: accept score: null and origin: "series_load" on the items of GET /watch/hits, and tell the alert apart by relation === "series": present it as a new series of your registration, not as a similar trademark, and do not assume score is a number when sorting or filtering.
2026-10-09 · [DATA] category is now also read from INAPI's record page: marks known only to the record page stop coming out without a category
INAPI's record page reports the registration category of each mark (trademark, Madrid Protocol mark, slogan…) and until today that datum was not read. A mark whose only source was the record page came out without category in GET /trademarks/{applicationNumber} ≡ get_trademark. The shape does not change and there are no new codes: the same seven.
- From today, every mark updated from the record page gets its
category(andcategoryLabel) if it had none. A category already present is not replaced. - Applied on 2026-10-10 (00:02 UTC), lot
categoria-detalle-261010. Marks updated before were filled with what the record page had already reported: 5,820 marks gained their category (792 of them Madrid Protocol); marks without a category went from 33,121 to 27,301. None doubtful or in conflict. - The remaining 27,301 had no stored detail: they were covered by the load of the legacy-import marks (entry of 2026-10-10 above), after which 1 mark remains without a category. Nothing is guessed.
- Gaining the category does not count as a change of the mark: it produces no watch notices and no new annotations.
GET /nice-classes/{niceClass}/stats≡get_class_stats: a mark that turns out to be of categorymarcamoves frompopulation.excludedUnknownCategorytopopulation.filedin each of its classes; one that turns out to be of another category leavesexcludedUnknownCategorywithout enteringfiled. Madrid Protocol marks were never counted in either figure: nothing changes for them.scopedoes not change: the record page does not report the protection scope.
What to do: nothing. If you keep a copy of the marks, read again the ones you hold without category; if you show excludedUnknownCategory, expect it to go down.
2026-10-09 · [DATA] Norm citations of the 6,450 annotations of 2024 with the full text: 5,010 added; 302 that the full text no longer supports, retired on 2026-10-10
Follow-up of the 2026-10-08 entry (6,450 M annotations of the 2024 Estado Diario now carry the full text of the bulletin). It affects the norm citations of those annotations: GET /annotations/{id}?expand=citations ≡ get_annotation, GET /trademarks/{n}?expand=norms ≡ get_trademark, the citedNorm, article and letter search filters, the citedNorms of the reports and the articles of GET /legal-norms ≡ list_legal_norms. The shape does not change.
- Done. 5,010 new citations in 2,118 annotations were read from the text that was missing (4,865 of Law 19.039, 125 of its Regulation, 19 of Law 19.880 and 1 of Law 18.575). The annotations of the group with at least one citation go from 6,044 to 6,279.
- Retired on 2026-10-10 (00:06 UTC), lot
normas-prosa-261010. The cut text produced some citations that the full text does not support: an «artículo 1» that was the beginning of an «artículo 19», an article 20 without the letter the full text does give, a citation without law whose law is now read. 302 citations in 262 annotations were retired, none of them among those that name their law expressly. The citations of the 6,450 annotations go from 23,664 to 23,362; one annotation is left with no citation. Trademarks, annotations and person links do not move. - No annotation or trademark is added or removed, and this produces no watch alerts.
What to do: nothing. If you counted or stored norm citations of 2024 M annotations, measure again from 2026-10-10.
2026-10-08 · [ADD] Series of a Madrid registration: baseApplicationNumber, seriesSuffix and relatedSeries; a series can be watched
An international registration (Madrid Protocol) can have more than one series in Chile. INAPI numbers each additional series with the number of the first one plus a suffix (991652209-2) and treats it as a separate application; the contract serves it as one more trademark, with its own registration number, classes and history. No additional series is loaded yet: this entry describes the shape they will arrive in.
baseApplicationNumberandseriesSuffix: only on an additional series, in the collection, search, similarity, batch, the trademarks of a party, the reports and the detail (GET /trademarks,POST /trademarks/search≡search_trademarks,GET /trademarks/{applicationNumber}≡get_trademark,GET /parties/{id}/trademarks≡list_party_trademarksand the reports).relatedSeries[]: on the detail of any series of a registration with several; it lists the others, withapplicationNumber,seriesSuffix(nullfor series 1),registrationNumber,status,classesandregisteredOn.- The three fields are omitted on any other trademark: its response does not change.
- In search, similarity, batch and portfolio the trademark appears once per series. In
GET /trademarks/renewal-calendar≡get_renewal_calendarthere is one line per registration: an additional series never gets a row of its own. POST /watch/applications≡watch_application(andDELETE≡unwatch_application, and theapplicationNumberfilter ofGET /watch/hits≡list_similarity_hits) accept, besides 1 to 12 digits, the number of a series (991652209-2). Any other shape is still a400 validation_error.
Details and examples: Series of a Madrid registration.
What to do: accept an application number with a hyphen and a suffix (letters or digits) wherever you require digits only today, and do not strip the suffix when normalizing: 991652209-2 and 991652209 are two different trademarks. If you show the detail, render relatedSeries as «this international registration has N series».
2026-10-08 · [BEH] Class statistics: coverage no longer declares a window with only the old import's bulletin
GET /nice-classes/{niceClass}/stats ≡ get_class_stats warned that, from 2024-04-19 to 2024-09-29, the corpus's only bulletin was the one of the old import. As of today the corpus has the Estado Diario of those days, so the notice is retired. No figure changes: the statistics were never computed differently inside that window.
coverage.windowsis always an empty list (it carried one item,legacy_only_bulletin_window). The field stays.coverage.notes[]no longer emits or declareslegacy_only_bulletin_window; five codes remain.coverage.annotationSources[]: the description ofestado_diarioreads "ingested since 2024-04-19" (was 2024-09-30) and the one oflegacy_mongono longer says it is the only bulletin of any period.- The note of
actsShare.oppositiondoes not change: opposition relations still exist only since 2024-09-30.
What to do: if you enumerated the coverage.notes[].code values, remove legacy_only_bulletin_window; if you displayed the notice of that window, stop displaying it.
2026-10-08 · [BEH] C11N and C11C gain a name; C11N joins the oposicion family
Two section codes that arrived with the 2024 Estado Diario had no name. They now carry the literal heading of the bulletin: C11N "Citaciones a oír sentencia en juicios de Nulidad" and C11C "Citaciones a oír sentencia en juicios de Caducidad". Their 24 annotations (23 and 1) no longer come out with sectionName: null, and GET /sections ≡ list_sections returns them with label in the three languages.
C11Ntakes theoposicionfamily, the one of its siblings of the nullity series: its 23 annotations match thefamily=oposicionfilter.C11Ckeepsfamily: null.- For that reason
actsShare.oppositionofget_class_statsmay rise by a few trademarks (at most 23 annotations across all classes) after the next recomputation. meta.dataAsOfofGET /sections≡list_sectionsmoves to2026-10-08T00:00:00.000Z.
What to do: nothing, unless you had stored the null name of those two codes.
2026-10-08 · [DATA] Estado Diario 2024: 6,450 annotations with cut text now carry the full text of the bulletin
The old import stored the observation of many 2024 Estado Diario annotations cut at about 2,000–2,500 characters. In production the text of 6,450 annotations of family M was replaced by the full text of the bulletin (observation in every annotation view: annotations[] of the record, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin).
- By section:
M32,770 ·M61,647 ·M111,468 ·M1259 ·M4232 ·M268 ·M76. Median length 2,186 → 4,353 characters. - For 2,069 the full text is new for the contract. For the rest it was already served through
observationDerived(it came from the record page sibling): thereobservationis now complete. - 24 annotations were left out on purpose (their record-page sibling would have stopped being comparable).
- Nothing else changes: no annotation is added or removed, and the
id, date and other columns keep their value. The new text is our bulletin reader's version, so some words the old import had glued together now come out separate. That night's chain adds norm citations from the longer text (about +5,010 expected, not yet measured).
Same shapes, no field added or removed.
What to do: if you stored or hashed the text of these annotations, read it again. Repeat any measurement that searched or counted words in 2024 M annotations.
2026-10-08 · [DATA] Estado Diario from 2024-04-19 to 2024-09-27: 681 acts the PDF reader was not reading are added
The PDF reader skipped some entries of the 2024 bulletins. With the reader fixed, 681 new annotations are added, almost all section D5A, over 11 days (annotations[] of the record, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, POST /annotations/search ≡ search_annotations, get_daily_bulletin): 2024-04-19 1 · 04-30 77 · 05-16 2 · 05-20 1 · 05-23 145 · 06-13 2 · 06-26 87 · 07-03 129 · 08-22 62 · 08-26 103 · 09-24 72. The 77 of 2024-04-30 came with the act fused to the trademark name; they are stored with the act separated (57 «Anotación de Cambio de nombre» and 20 «Anotación de Transferencia total»).
- Annotations of the period: 438,815 → 439,496, no duplicates. No existing annotation is completed and no trademark is created.
- Not included: the 354 entries of the type-5 bulletin filed under 2024-05-21, a Chilean holiday. The file has no printed date and has 53 pages against the usual 17–19; the hypothesis that it accumulates the type-5 bulletins of 16, 17 and 20 May is not confirmed. They wait for INAPI to clarify the notification date.
⚠ The new rows were created on 2026-10-08, the date detectedAfter uses, so the 681 old acts show as detected today although their date is from 2024.
Same shapes, no field added or removed.
What to do: also filter by date or ignore the 2026-10-08 batch. Repeat any measurement that counted 2024 annotations of section D5A or of those days.
2026-10-08 · [DATA] Estado Diario from 2024-04-19 to 2024-09-27: 927 more «second acts» are added, and the 2024 work is complete
A «second act» is a bulletin entry on the same trademark, day and section as another one, but saying something different (for example a second resolution on the same day). They were left out of the first batch on purpose; they are now in.
- 927 new annotations, all section
M11, spread over 106 of the 108 business days (annotations[]of the record,GET /trademarks/{n}/annotations≡list_trademark_annotations,POST /annotations/search≡search_annotations,get_daily_bulletin). - 4 entries are not inserted (all from 2024-06-17,
M11): three are the same resolution as a row that already exists, and one is a different act whose text the PDF reader garbled. Their review is still open. - In total, the 2024 archive work: 45,672 new annotations (44,745 + 927) and 69,532 existing ones completed. Annotations of the period: 393,143 → 438,815, no duplicates. No trademark is created.
⚠ Same as the first batch: the new rows were created on 2026-10-08, the date detectedAfter uses, so they show as newly detected although their date is from 2024.
Same shapes, no field added or removed.
What to do: also filter by date or ignore the 2026-10-08 batch. Repeat any measurement that counted 2024 annotations of section M11.
2026-10-08 · [DATA] Estado Diario from 2024-04-19 to 2024-09-27: 44,745 missing annotations are added and 69,532 existing ones gain their parties
The corpus held the Estado Diario of those 108 business days only as the old import left it: family M alone, with representante, solicitante and tipo_resolucion empty. It now also has the bulletins themselves, read from a public archive (213 PDFs).
- 44,745 new annotations (
annotations[]of the record,GET /trademarks/{n}/annotations≡list_trademark_annotations,POST /annotations/search≡search_annotations,get_daily_bulletin): familyA20,330, familyC12,721,D5A10,901, familyM793. Before, the corpus had no row of familyA,CorD5Aof the Estado Diario in that period. - 69,532 existing family-
Mannotations gain their parties (representante,solicitante) andtipo_resolucion, which they had empty. Their text,idand date do not change. - Annotations of the period: 393,143 → 437,888, with no duplicates. No trademark is created.
- Not included: the type-5 bulletin of 2024-05-29 (INAPI published the 2024-04-29 one there), the
D5Aacts of 16, 17 and 20 May 2024 (that bulletin is not in the archive), 14 entries of trademarks that are not in the corpus, 78 acts the reader cannot resolve, and the «second acts» of the same day, which came later in a batch of their own (see the entry above).
⚠ The new rows were created on 2026-10-08, and that is the date the detection filter uses: POST /annotations/search ≡ search_annotations with detectedAfter returns them as newly detected even though their date is from 2024.
The shape does not change: no field is added or removed.
What to do: if you treat every detectedAfter hit as news, also filter by date or ignore the 2026-10-08 batch (~45,000 old acts). Repeat any measurement that counted 2024 annotations by family or section.
2026-10-07 · [BEH] Boolean parameters only accept true or false: any other value is a 400 validation_error
Until today, liveOnly, allClasses and transitional read any value other than true or 1 as false: GET /trademarks?liveOnly=yes answered 200 with the unfiltered result and no warning, although the reference declares the parameter as a boolean. From today a value that is not exactly true or false (lowercase) answers 400 validation_error, with the parameter in errors[].field, like every other invalid parameter. That includes yes, si, TRUE, the empty value (?liveOnly=), a repeated parameter and also 1 and 0, which used to be accepted.
liveOnly:GET /trademarks,POST /trademarks/search≡search_trademarks;GET /trademarks/renewal-calendar≡get_renewal_calendar;GET /insights/trademarks,POST /insights/trademarks/search≡get_trademark_insights.allClasses:GET /trademarks,POST /trademarks/search≡search_trademarks.transitional:GET /legal-norms/{id}/articles/{article}.
In a JSON body and over MCP the value is a JSON boolean (true / false); a text such as "yes" or a number, which used to be read as false, is now rejected. true, false and the absent parameter return exactly what they returned before. dedupe already behaved this way and does not change.
What to do: if you send liveOnly=1, allClasses=1 or transitional=1, change it to true; if you send 0, to false, or drop the parameter. If you were sending any other value, the filter was not being applied: review the results you obtained with it.
2026-10-07 · [BEH] Sections: D5A are filed recordal requests, W088 and W635 carry INAPI's name, and renewals[] no longer counts D5A
Values change in the section catalogue (GET /sections ≡ list_sections) and in the annotation fields that quote it (sectionName, sectionLabel, sectionFamily). The shape does not change.
D5Agoes from «Anotaciones inscritas (segunda tabla del boletín de altas)» to «Anotaciones presentadas (segunda tabla del boletín de altas)»; the English label is now "Recordal requests filed (second table of the new-entries bulletin)". That Estado Diario table lists the recordal requests filed that day: the recordal itself comes later (a median of about 34 days) or never (some are rejected or abandoned).renewals[]andlastRenewalOn(GET /trademarks/{applicationNumber}≡get_trademark) no longer count anyD5Arow. Until today onlyD5A+ARENTLTwas left out; with other renewal acts (AREPTLT,AREN,AREP) the row was published as a renewal. The rows stay inannotations[].W088: «Aceptase Tramitación (Publicación)» → «Por Cumplido lo Ordenado»;sectionFamilygoes frompublicaciontonull.W635: «Cúmplase» → «Por interpuesto recurso de apelacion.-»;sectionFamilygoes fromresoluciontonull. These are the names INAPI publishes; the previous ones were deduced. It affects annotations from 2002 to 2012.W160«Rechaza Resolucion (Plazo 15 Apelacion)» andW148«Fallo Notif. Rechazado (15 dias Apela)»:sectionFamilygoes fromnulltoresolucion, the family of their current equivalents024and486. They are the rejections issued before February 2012; thefamily=resolucionfilter now returns them.meta.dataAsOfofGET /sectionsmoves to 2026-10-07.
Class statistics do not change: no W code joins the fondo family.
What to do: if you stored section names or families, read GET /sections again. If you showed a D5A row as a recorded annotation or as a completed renewal, show it as filed; the completed renewal is in renewals[].
2026-10-07 · [FIX] MCP: the weekly_watch prompt pages while hasMore is true, and tool descriptions say what the tools do
The weekly_watch prompt asked to walk list_similarity_hits "until next is null"; since today that list always returns next, so the prompt never ended. It now pages while hasMore is true. availability_check and the search_trademarks description name the real channels of the similarTo similarity engine (phonetic, lexical, semantic and distinctive element) and point to matchedBy. get_party, get_legal_norm, get_section, get_system_freshness and list_sync_runs state the permission they require and what they return for an unknown identifier. No change to inputs, responses or the REST API.
What to do: nothing. If you copied the weekly_watch prompt into your own agent, change its stop condition to hasMore.
2026-10-07 · [FIX] Sub-processors: Supabase is only involved in the Tarno application
The sub-processors list and the data processing agreement described Supabase as the store for watching and batch jobs. Watching is kept in the Service's database and batch jobs no longer exist. Supabase remains a sub-processor only for the Tarno application (user accounts and application data). The DPA no longer lists batch jobs or their 30-day deletion period.
What to do: nothing. The data your organisation writes through the API and the MCP does not go to Supabase.
2026-10-07 · [BEH] GET /trademarks/{applicationNumber}/deadlines ≡ get_deadlines: the deadline to appeal a partial acceptance (444) is now published
Until today the appeal deadline was only given for a rejection (024); a mark whose most recent resolutive movement was a partial acceptance (444) came out in omitted[] with partial_acceptance_not_published. Now the 444 anchors like the 024: the 15th business day strictly after its publication in the Estado Diario, with the same three norms (Ley 19.039 art. 17 bis B and art. 11, Ley 19.880 art. 25). INAPI counts the term from the Estado Diario, and its Res. Ex. 138/2022 says the e-mail is a courtesy notice that does not replace the legal notification.
deadlines[].anchor.sectionCodecan be"024"or"444"(before, always"024").anchor.kindis stillrejection_notice: in a partial acceptance the appeal is against the part of the coverage that was not accepted.- Between a
024and a444the most recent one anchors; on the same date, the024. If resolutions that do not anchor come later, the date is computed from the most recent of the two, withcaveat: later_resolution_on_record. omitted[].reasonno longer emits or declarespartial_acceptance_not_published; four values remain. A Madrid designation, an anchor on a non-business day or a span outside the declared calendar are still omitted with their reason, for a444too; no estimate is ever given.- The
appeal_closemilestone ofmilestones[]now also appears on marks whose most recent resolutive movement is a444, with the same shape. On a watched application inrechazadaordenegadawith a444as its most recent movement,watchEndsOnbecomes that date (watchEndBasis: "appeal_close").
What to do: accept "444" in anchor.sectionCode; if you enumerated the omitted[].reason values, remove partial_acceptance_not_published; and expect the appeal_close milestone on marks with a partial acceptance.
2026-10-07 · [BEH] POST /watch/applications ≡ watch_application: the plan cap is the organisation's, the same for all its keys
The cap of active watched applications (watch_limit_reached, limit) is now your organisation's, the same unit the watched applications were always counted over. Before, it was read from the key making the request: with two keys of one organisation and different caps, one could add an application while the other got 422 watch_limit_reached with the same count. Now every key of the organisation sees the same limit. Same shape, codes and statuses; without a contracted cap, limit is still 500.
What to do: nothing. If you spread watched applications across keys of one organisation, count on a single cap for all of them.
2026-10-07 · [ADD] Legal documents published: terms of service, personal data policy, data processing agreement (DPA) and sub-processors
New Legal section in this documentation, current version dated 2026-10-07: the terms of service (what the service is and is not, keys, permissions, acceptable use, suspension and termination), the corpus personal data policy (which personal data it holds, which INAPI source it comes from, what it is processed for, how long it is kept and how to exercise your rights), the data processing agreement for the data your organisation writes into the service, and the list of sub-processors. The API contract does not change.
What to do: read them. In particular terms § 7.3: personal data you obtain from the service is used only to identify the holder or representative of a trademark and is not disclosed to third parties beyond that purpose.
2026-10-07 · [ADD] SLA published: 99.5% monthly availability, 24 h RPO, 4 h RTO and an announced maintenance window
The Service level page is no longer best-effort: it is now the SLA of the REST API and the MCP. 99.5% monthly availability and under 0.5% 5xx responses; maximum data loss (RPO) of 24 h, and 1 h for API keys and organisations; recovery time (RTO) of 4 h within support hours; maintenance only in the declared window (Sundays 02:00–06:00, America/Santiago, at most 4 h per month), announced 48 h ahead. The daily sync is not maintenance: it does not interrupt the service. The page explains how each figure is measured, the exclusions (force majeure, INAPI unavailability, use beyond quota) and the p95 latency objectives. The API contract does not change.
What to do: nothing. If your integration assumed "no numeric guarantee", you can now rely on these figures.
2026-10-07 · [BEH] GET /sections ≡ list_sections: meta.dataAsOf gives the day the INAPI code legend was last read
Until today it was the only catalog with meta.dataAsOf: null. Now it is the day on which the section-code legend served by the catalog was last read in an INAPI publication (search engine responses, Estado Diario bulletins, digital file index), given as the start of that day in UTC (today 2026-10-02T00:00:00.000Z). It is a day, not an instant: those readings are dated by day, and the start of the day is never later than the real reading. meta.coverageThrough stays null. Same shape.
What to do: nothing. If you treated the null of /sections as "no date", you can now show the date.
2026-10-07 · [FIX] POST /annotations/search ≡ search_annotations and the sync.finished webhook: with dedupe (default), a repeated act is no longer served as a new act
With dedupe, the rows that describe the same act are presented once, with the detection date of the first one. That grouping is computed after each daily sync, and in the interval between the end of the sync and the grouping a repeated row could appear in "changes since" as a new act (the late copy of an act you had already received, or the two rows of that night's act as two acts), and then stop appearing. Now, with dedupe, detectedThrough does not go past the last completed grouping: what comes after is served as soon as it is grouped. The webhook announces the same horizon. Same shape; detectedThrough can only be equal to or earlier than before. With dedupe: false nothing changes.
What to do: nothing. Keep asking with detectedAfter = the previous detectedThrough. If you stored acts received in that interval, there may be copies of the same act: fetch them again by id (GET /annotations/{id}) and group them by eventId ?? id.
2026-10-07 · [BEH] Catalogs: meta.dataAsOf gives the date of their source and GET /holidays ≡ list_holidays gains meta.coverageThrough
Same shape; two envelope values change on GET /nice-classes ≡ list_nice_classes, GET /legal-norms ≡ list_legal_norms, GET /holidays ≡ list_holidays and GET /sections ≡ list_sections. meta.dataAsOf is the last successful load or check of each catalog's own source, never the time of your request: for the Nice classes, the last check of the headings against WIPO (was null); for the legal norms, the last sync with BCN (was null); for the holidays, the last load of the calendar (was the last update of the trademark corpus). For the sections it stays null: their load date is not recorded. meta.coverageThrough stays null on catalogs (they have no "through"), except on holidays: the last day with a declared calendar (today 2030-12-31, the same value as meta.coverage.to); deadlines ending on or before that date are counted with declared holidays. Catalogs are served with a cache of up to one hour.
What to do: nothing. If you showed the holidays' dataAsOf as the corpus freshness, use the one of a trademark collection.
2026-10-07 · [FIX] POST /annotations/search ≡ search_annotations and the sync.finished webhook: detectedThrough no longer gets ahead of acts still being loaded
detectedThrough is the safe horizon: asking next with detectedAfter = that value must not skip anything. With some corpus loads that do not go through the daily sync, the horizon could get ahead of acts not yet visible which, once committed, fell before it: those acts never showed up in «changes since». The horizon now also waits for those loads. Same shape; detectedThrough can only be equal to or earlier than before.
What to do: nothing. Keep requesting with detectedAfter = the previous detectedThrough.
2026-10-07 · [FIX] Webhook sync.finished: days missed while the sender was down are sent when it returns, and a re-run chain corrects the pending body
Three fixes to the sending, with the same body shape and headers. A fault on the sender could leave a subscriber without the day's notice: the queue no longer stalls, and a notice that does not arrive is detected and reported to the operator. If the sender was down for more than 24 h, the days that did not go out are sent when it returns (up to 7 days back), each with its own corpusDate. If the day's chain is re-run before your delivery completes, the retries carry the new body (same id). And a delivery you already received can no longer arrive again because recording its success failed for a moment.
What to do: nothing if you deduplicate by id; process notices in corpusDate order, not in arrival order.
2026-10-07 · [FIX] MCP over HTTP: when the key cannot be verified, the answer is the same 503 service_unavailable as REST
During a transient outage of the service that validates keys, and for a key the server had not seen in the last minutes, REST answered 503 service_unavailable while the MCP could still serve the call, without counting it against the monthly quota or the burst limit. The MCP now answers what REST answers: HTTP 503 with the usual error body (code: "service_unavailable", the same requestId as the X-Correlation-ID header), without running the tool. Any other failure to verify the key is the neutral 500 internal_error, as on REST.
What to do: nothing, if you already retry REST 503s: treat an MCP HTTP 503 the same way (retry with backoff).
2026-10-07 · [BREAK] GET /watch/hits ≡ list_similarity_hits: the list is read in the order your organisation came to see each match, next is always returned and earlier cursors start over; new field visibleSince
The list used to be paged by the match id and ended with next: null. A client that kept "the highest id read" silently lost the matches that became readable to it later with an old id: those of an application it subscribed to when another organisation already watched it, and those that reappear when it stops watching one of its own trademarks.
- Order: ascending
visibleSince, thenid.visibleSinceis new on every match: since when your organisation can read it (ISO instant in UTC, with microseconds) — the latest of its detection, the start of your current subscription to the watched application and the last time you stopped watching the matched application. nextis always returned, on the last page and on an empty list too. The end of a read ishasMore: false. Keep the lastnextand send it later ascursor: you get only what you came to see since.- A match can reach you again with its same
id(it reappears, or you unsubscribed and subscribed again). One that was closed and comes back arrives with a newid, as before. - A cursor belongs to its filter: the one obtained with
applicationNumber=Nonly continues that read; with another filter the list is read from the beginning. - Cursors issued before this change are not recognised: the read starts from the beginning. It is not an error.
- Guarantee: the list only serves what is settled, so a match that finishes being stored late is not left behind your cursor: it is delayed (milliseconds; seconds during a maintenance task), not lost. It does not cover an operation of the service longer than 60 seconds, a manual data correction or a change of the
scorethreshold. A read with nocursoralways walks everything.
GET /watch/applications ≡ list_watched_applications does not change.
What to do: (1) page while hasMore is true — a loop "until next is null" no longer ends; (2) keep the last next and use it as cursor on the next read, instead of keeping the highest id and re-reading the list; (3) insert or update by id; (4) drop the cursors stored before today; (5) keep one cursor per filter you use. Guide: Similarity watch.
2026-10-07 · [BEH] Watch (GET /watch/applications ≡ list_watched_applications): watchEndsOn is never before the day the application was seen terminal
When a watched application became terminal with its last day of watch already past (an expiry published late, a rejection older than its margin), the watch ended at that same check and watchEndsOn was never readable. That last day is now, at the earliest, the day the service saw it terminal: the watch is announced that day and ends the next. The fields and their possible values do not change.
What to do: nothing.
2026-10-07 · [BEH] Watch (GET /watch/hits ≡ list_similarity_hits, GET /watch/applications ≡ list_watched_applications): an application missing from the corpus for a single day no longer closes anything
When an application is no longer in the corpus, its watch no longer ends, and its matches no longer stop being listed, at the first daily check that does not find it: the next day's check must not find it either. It therefore takes between one and two days instead of up to one; meanwhile the match is still listed, with matched: null. A match that stopped being listed for that reason comes back — with a new id, like any other match that comes back — if its application reappears and is live. The shape of the responses does not change.
What to do: nothing.
2026-10-07 · [FIX] OpenAPI and MCP: the contract version moves to 2026-10-04, the title is «Tarno Read-API» and the MCP server is named tarno-read-mcp
Only the published text changes: no field, route, tool, scope or code changes. The OpenAPI info.version, the MCP server version and the manifest's contractVersion said 2026-09-29 although the changes of October 1–4 were already in force; they now say 2026-10-04. The OpenAPI is titled «Tarno Read-API», and the MCP initialize answer names the server tarno-read-mcp, as the MCP guide always said. GET /watch/applications, DELETE /watch/applications/{applicationNumber}, GET /watch/hits and GET /system/me gain in the OpenAPI the description their MCP twins already had.
What to do: nothing, unless you compare the contract version or the MCP server name: update the expected value.
2026-10-04 · [FIX] GET /trademarks ≡ search_trademarks (with q or similarTo): the lexical channels work again with liveOnly, status=caducado and the expiry, milestone and renewal filters
From 2026-10-03, with liveOnly=true, status=caducado, expiresFrom/expiresTo, milestoneFrom/milestoneTo (milestoneKind expiry, renewal_window_open, grace_end or no milestoneKind) or renewalStatus, the lexical channels (exact, fuzzy) did not run: the answer came only from the phonetic and semantic channels. They now run again: the text engine scores the matches of q and the filter is applied afterwards. With those filters, results with matchedBy exact/fuzzy come back; in particular similarTo again finds marks that differ only by a typo or a suffix. With any other filter, same list and same order.
The first search of a watched application no longer ends done when a channel failed: it is retried, and baselineStatus stays pending meanwhile.
What to do: nothing.
2026-10-04 · [BEH] Watch: matches move by themselves (daily cross, closing by status) and WatchedApplication gains terminalSince, watchEndsOn and watchEndBasis
Same four operations, inputs, scopes and codes; what changes is what you see over time in GET /watch/hits ≡ list_similarity_hits and GET /watch/applications ≡ list_watched_applications:
- After each complete sync, every watched application is crossed with the live applications that changed: new matches arrive with
origin: "sync". - A match stops being listed if it becomes terminal (checked once a day) or no longer reaches the threshold; if it comes back, it appears as a new row (new
idanddetectedAt). baselineStatuscan go back topendingif the watched application's denomination or classes change.- If the watched application becomes terminal, it is still watched until
watchEndsOn; the day after, it disappears from the list, its matches stop showing and it stops counting for the plan cap (DELETEanswers404; a newPOST,422 application_terminal).
New fields, null while the application is live: terminalSince (ISO instant), watchEndsOn (YYYY-MM-DD, last day of the watch; not a legal deadline) and watchEndBasis (appeal_close, grace_end, rejection_margin or expiry_margin).
What to do: re-read matches by cursor and store the highest id read; do not assume a listed match stays there. Optionally, show watchEndsOn.
2026-10-04 · [BREAK] Watch: the ten watchlist operations are retired; four operations watch an application by its number
Retired, on REST and MCP and with no coexistence period: POST/GET /watchlists, GET/PATCH/DELETE /watchlists/:id, POST/GET /watchlists/:id/items, DELETE /watchlists/:id/items/:itemId, GET /watchlists/:id/hits and GET /watchlists/:id/deliveries (≡ create_watchlist, list_watchlists, get_watchlist, update_watchlist, delete_watchlist, add_watchlist_item, list_watchlist_items, remove_watchlist_item, list_watchlist_hits, list_watchlist_deliveries). Gone with them: items by denomination, per-client filters, the per-list webhook with X-Tarno-Signature, Idempotency-Key, and the codes watchlist_too_large, watchlist_retired, idempotency_conflict and concurrent_modification.
They are replaced by (scopes watch:read/watch:write; the key must have an organisation):
POST /watch/applications≡watch_application— body{ applicationNumber };201(or200if already watched) with aWatchedApplication.DELETE /watch/applications/{applicationNumber}≡unwatch_application—204.GET /watch/applications≡list_watched_applicationsandGET /watch/hits≡list_similarity_hits— common envelope withcursorandlimit(50 by default, 200 at most).
New codes: 422 application_terminal (with applicationStatus), 422 watch_limit_reached (with limit) and 429 watch_daily_limit_reached (with limit and resetsAt, no Retry-After). The MCP prompt weekly_watch takes applicationNumber instead of watchlistId. There is no webhook and no per-client parameters; existing lists are not migrated.
What to do: stop calling /watchlists; subscribe each application with POST /watch/applications and read matches with GET /watch/hits. Full details in the Watch guide.
2026-10-04 · [ADD] POST /annotations/search ≡ search_annotations: detectedAfter alone is the corpus-wide feed of acts, and every annotation carries denomination
detectedAfter (with an optional detectedUntil) no longer needs a trademark: without applicationNumber, applicationNumbers, partyId or taxId it returns every act the corpus learned of since that instant, in detection order, by cursor and with meta.detectedThrough. In that mode only detectedUntil, dedupe, limit and cursor are accepted; any other filter is a 400 naming it. Before, { detectedAfter } alone was a 400.
Every Annotation (search, GET /trademarks/{n}/annotations ≡ list_trademark_annotations, GET /annotations/{id} ≡ get_annotation and the daily bulletin) gains denomination, the trademark's denomination (empty string if it has none). Also, a sweep with a fixed detectedUntil is now frozen: rows stored afterwards take no part, not even in the dedupe collapse.
What to do: nothing; if you follow thousands of marks, you can read the feed and filter on your side instead of asking in batches.
2026-10-04 · [ADD] Webhook sync.finished: a signed notice when the day's corpus is ready
When the daily sync finishes, Tarno sends one POST per day to each subscriber, with X-Tarno-Event: sync.finished and the signature X-Tarno-Signature: t=<unix>,v1=<HMAC-SHA256 of "t.rawBody"> (±300 s tolerance). The body (version 1) carries id (sync.finished:<date>, for idempotency), corpusDate, ok, syncExitCode, dataAsOf, coverageThrough, detectedThrough, watchWatermark and per-source counts; no client data. It is sent even when a source failed (ok: false). Retries at 1 min, 5 min, 15 min, 1 h and 4 h.
The subscription (URL and signing secret) is configured by Tarno for now; there is no REST route or MCP tool for it. No existing route changes.
What to do: if you want the notice, ask for a subscription and verify the signature; you can stop polling for freshness.
2026-10-04 · [BEH] POST /watch/applications ≡ watch_application: may answer 503 service_unavailable if the plan cap cannot be read
The plan cap of watched applications (watch_limit_reached, limit) is now read from a different source. The value, shape and codes do not change. The only new outcome: if that source does not answer while an application is being added, the call fails with 503 service_unavailable (retryable) on both transports, instead of an internal error. Listing, unwatching and reading matches are unaffected.
What to do: treat that 503 like any other: retry later.
2026-10-04 · [ADD] Webhook sync.finished: test deliveries marked X-Tarno-Test: true
Tarno can send a subscriber a test delivery on demand. It carries the latest real notice byte for byte (same id, so an id-deduplicating receiver skips it) or, if there has been no notice yet, a synthetic body with type, version, an id starting with sync.finished:test: and "test": true. It is signed like any delivery, adds the header X-Tarno-Test: true and its own X-Tarno-Delivery (test-<uuid>), is a single attempt and does not replace the day's delivery.
What to do: if you receive X-Tarno-Test: true, verify the signature, answer 2xx and do nothing else.
2026-10-04 · [ADD] GET /system/me ≡ get_system_me: usageHistory, the key's accounted usage over the last 12 months
New field usageHistory, always present and with the same JSON on both transports: windowMonths (12), asOf (last recorded usage of the key, or null) and months, newest first, one entry per UTC month with recorded usage: month (YYYY-MM), requests, quotaUnits, rejectedRequests (429), logoAnalyses and byScope (scope, requests, quotaUnits, rejectedRequests). It covers only the presented key, not other keys of the organisation.
A missing month means no recorded usage (accounting starts on 2026-10-03), never a computed zero. The current month may lag quota.consumed by about a minute.
What to do: nothing; use it if you want to show the key's historical usage.
2026-10-02 · [BEH] expiresOn, milestones[] and status: a Madrid mark follows the renewal (MAD02) and extinction (MAD01) notified by WIPO
A registered Madrid mark (9-digit applicationNumber) with a MAD02 from one year before INAPI's expiry onwards and no later MAD01 now publishes expiresOn = INAPI's expiry + 10 years per cycle (Madrid Protocol, art. 7.1). Its expiry milestone becomes basis: "derived", with anchor inapi_expiry, rule: "10_years_after_previous_expiry_on_wipo_renewal" and legalBasis art. 7.1. Expiry filters and groupings use the published date; sort=expiresOn still orders by INAPI's date.
If its latest MAD01 falls on or after the expiry and after the latest MAD02, it publishes status: "caducado", statusInapi: "registrada", statusStale: true; filters, facets and liveOnly follow it. 157 marks move their expiry and 61 become caducado. Same shapes.
What to do: nothing.
2026-10-02 · [BREAK] encumbrances[] (GET /trademarks/{applicationNumber}?expand=encumbrances ≡ get_trademark): a modern charge rules only once INAPI registers it; new state rejected
A modern A0xI annotation is the filing of a charge, not its registration. Until now every one came out active (or released); now each is resolved through INAPI's resolutions:
- registered (
1104,638, or639for a release) →active/released, as before; rejected(new value) — rejected (537); never in force;abandoned—536,637or927; never in force;filed— no closing resolution yet.
A release (A0xA) lifts only once registered. A charge that never ruled carries filedOn (its A0xI date) and recordedOn: null; one in force or released keeps recordedOn and filedOn: null. None of filed, rejected, abandoned counts as a charge in force or should trigger alerts. Of 4,691 modern charges, active goes from 2,267 to 1,677; 186 rejected, 464 abandoned and 1 filed appear. 312 marks are left with none in force.
What to do: an exhaustive switch on state must handle rejected; count only active as in force.
2026-10-02 · [DATA] Section 1076 is «Resolución de aceptación e inscripción de renovación TLT» and now counts as a renewal
Code 1076 carried the name of OBS03 («Escrito de anotación - Cumplimiento de observaciones»). It is corrected: its 87,483 annotations (87,349 marks) now carry sectionName «Resolución de aceptación e inscripción de renovación TLT», their own sectionLabel and sectionFamily renovacion instead of presentacion, so the sections listing by family and the family=renewal filter include it.
In the trademark report (renewals[], lastRenewalOn), 1076 joins the renewal acts: 2 marks gain their first renewal and for 54,860 lastRenewalOn moves to the date the renewal was entered. The same act no longer comes out twice (also 1077 with A3 + AREPTLT). renewalStatus, expiresOn and the milestones do not change. Same shapes.
What to do: nothing.
2026-10-02 · [FIX] status and registrationNumber: a grant invalidated afterwards (Ley 19.880) no longer derives registrada
The derived status promoted a mark in a pre-grant state to registrada/concedida when there was a witness of the grant (registration number, grant annotation or a C class), without looking at a later invalidation. Now, if the latest invalidation annotation (957 or 957B) is after the latest grant annotation, status is INAPI's value, statusInapi is null and statusStale is false; also registrationNumber is null in lists and omitted in the report. A new grant afterwards brings the derivation back. The status= filter, groupBy=status, the facets and sort=status follow the same rule. Today it affects one mark. Same shapes.
What to do: nothing.
2026-10-02 · [DATA] Estado Diario D5A annotations: each act moves to the mark whose registration number the bulletin prints
Section D5A («Anotaciones inscritas») prints the mark's registration number; since 2026-09-03 it was stored as the application number, so every D5A annotation sat on another mark. The right mark is now resolved:
annotations[](report,list_trademark_annotations,search_annotations,get_daily_bulletin): 41,375 rows move to their mark, with the sameid, date and text. 37,187 marks lose acts that were never theirs and 37,182 gain theirs.- 6 rows whose registration number no mark holds are removed: their
idanswers404. renewals[]/lastRenewalOn, transfers (A07T) and name changes (A01) underD5Amove with their rows.D5A+ARENTLTstays out ofrenewals[]for now.
Same shapes.
What to do: if you stored D5A annotations per mark, read them again.
2026-10-02 · [BREAK] encumbrances[]: charges from before 2012, two new state values and filedOn
Charges filed before INAPI's 2012 code change (W340 medida precautoria, W360 prenda, W370 prohibición) now appear in encumbrances[], with registrationSection set to the old code and the creditor in beneficiary.
stategainsfiled(filed, not registered) andabandoned(accepted, then lapsed for non-payment; never in force).- New field
filedOnon every entry: the filing date of an old charge;nullon modern ones. - A modern release lifts the old charge of its kind when only one is open. With two or more, the release is listed as an orphan (
released,recordedOn: null) and those charges stayactivewithreleaseObservable: false.
The answer changes for 88 marks; all others only gain filedOn: null.
What to do: an exhaustive switch on state must handle filed and abandoned; neither is a charge in force.
2026-10-02 · [BEH] 128 section codes gain INAPI's official name: sectionName/sectionLabel and /sections ≡ list_sections/get_section
128 codes that had no name (8,130 annotations, almost all old) take the one INAPI's Expediente Digital index gives them; among them A08I/A08A «Anotación de Quiebra», A09A, A03M/A05M and the 705A-D/773A-D families. Those annotations go from sectionName: null to INAPI's literal; sectionLabel always carries es-CL, and en/pt-BR for 112 of the 128. The catalogue now has 674 codes, 671 with a name. Same shapes; catalogues (/sections) may be served from a cache of up to one day.
What to do: nothing.
2026-10-02 · [BEH] GET /trademarks?similarTo= ≡ search_trademarks: pages of a similarity search no longer repeat or skip marks
The distinctive channel depends on a generic-term detector whose result could vary between pages, so the cursor repeated or skipped marks. The next cursor now carries what the channel did on the first page (the distinctive core it searched, or that it did not run), and every later page walks the same candidate set. The cursor stays opaque and grows slightly; earlier cursors keep working. If the detector fails on a later page, meta.genericTerms, distinctiveCore and decompounded are omitted on that page.
What to do: nothing; treat the cursor as opaque.
2026-10-02 · [BEH] POST /logos/analyze ≡ analyze_logo: a 429 logo_global_quota_exceeded no longer spends the key's daily logo quota
When the service-wide daily cap of logo analyses is full, the call is rejected with 429 logo_global_quota_exceeded as before, but the unit taken from the key's own daily logo quota (20 by default) is given back. Before, retrying on a full day could exhaust the key's own quota without a single analysis. Same shapes, codes and Retry-After. The general monthly quota is still taken before the tool runs, including on a 503 logo_analysis_not_configured.
What to do: nothing.
2026-10-01 · [BEH] PATCH /watchlists/:id ≡ update_watchlist: a retired watchlist can no longer be edited or reactivated → 409 watchlist_retired
A watchlist retired with DELETE /watchlists/:id (delete_watchlist) became terminal: any PATCH/update_watchlist on it, in particular {"status":"active"} or "paused", answered 409 with the code watchlist_retired and wrote nothing. To watch again, a new watchlist had to be created. GET, hits and deliveries of a retired one were unchanged. Watchlists were retired entirely on 2026-10-04.
What to do: nothing; use watching by application.
2026-10-01 · [BEH] The party filters holder, representative and taxId accept at most 20 values of up to 256 characters
The three party filters were unbounded lists. Now, in GET /trademarks, POST /trademarks/search, GET /insights/trademarks, POST /insights/trademarks/search, GET /trademarks/renewal-calendar and their MCP twins (search_trademarks, get_trademark_insights, get_renewal_calendar), each accepts up to 20 values (since 2026-10-16, several values are sent by repeating the parameter: a comma no longer separates names) and each value up to 256 characters, the same limit as q.
Exceeding the limit is a 400 validation_error with errors[].code = "too_big" on the field (over MCP, the equivalent input error; a value that is too long is reported as holder.0, holder.1…). With 20 values or fewer nothing changes.
What to do: nothing if you send one value per filter. If you send the party list of a portfolio, split it into requests of 20.
2026-10-01 · [DATA] GET /system/coverage ≡ get_system_coverage: acts no longer publishes false zeros
acts was 0 on days whose bulletin is in the corpus: the tipo2 bulletins from 2026-08-26 to 2026-09-08, several rows of 2026-08-26 and the altas bulletins before 2026-09-09 (between 950 and 1,615 acts per day). Each row now starts with the acts of that bulletin already in the corpus, and the historical rows have been corrected: the values for those dates go from 0 to their actual count. The shape does not change.
Acts that arrive later through paths that do not record the day are still not added: acts may fall short, never overcount. The same value is read in meta.bulletins[].acts of GET /reports/daily-bulletin ≡ get_daily_bulletin.
What to do: nothing. If you stored acts for those dates, read it again.
2026-10-01 · [BEH] A search whose channels all fail returns 503 service_unavailable, not an empty 200
Affects GET /trademarks and POST /trademarks/search ≡ search_trademarks when they carry text (q or denomination). The search runs several channels (exact text, fuzzy text, phonetic and, when available, semantic); a failing channel is dropped and the others continue. Until now, if all of them failed, the response was 200 with results: [], a false "no trademark matches". It is now 503 service_unavailable (application/problem+json, retryable), the same over REST and MCP.
If some channels fail but not all, nothing changes. A channel with no matches is not a failure: a search without results is still an empty 200. Similarity search (similarTo, similarToTrademark, oppositionWindow: "open") and the reports that use it already returned an error in this case. No shape changes: 503 was already declared in the OpenAPI.
What to do: treat a search 503 service_unavailable as "try again", not as "no matches".
September 2026
2026-09-30 · [BEH] MCP: explain_section_code returns content (previously always no_se_sabe)
MCP only. Until now the tool answered known: false / no_se_sabe for every code. It now explains 23 codes (012, 024, 444, 484, 485, 486, 576, 662B, 664B, 911, 912, A09I, E1, M6, M7, M8, M11, REC02, REC03, REC08, W148, W160, W230).
The response gains fields: explanation (markdown), title, status (measured · confirmed · contradictory · hypothesis), lastReviewed (YYYY-MM-DD, the date of the information), sources ([{ source, date }]) and knowledgeRevision ({ commit, date }). For an unknown code the new fields are null (sources: []). code, known and note keep their meaning; a code without an entry is still no_se_sabe, and a malformed code is 400 validation_error.
What to do: show lastReviewed next to the explanation and take status into account: hypothesis or contradictory are not settled facts.
2026-09-30 · [BEH] The MCP resources iris://catalogs/* require a key and scope; the holidays resource is no longer empty
resources/read now requires the same key and scope as the twin REST route: brands:read for iris://catalogs/states, iris://catalogs/sections and iris://catalogs/nice-classes (≡ GET /sections, GET /nice-classes) and insights:read for iris://catalogs/holidays/{year} (≡ GET /holidays). Without a key it returns the same authentication error as the tools; without the scope, the permission error. resources/list stays open, like tools/list.
In addition, iris://catalogs/holidays/{year} returned holidays: [] for every year; it now returns the holidays of the requested year, with the same declared coverage.
What to do: send your key when reading MCP resources too, with the matching scope.
2026-09-30 · [BREAK] formerParties[].retiredBy changes from a string to an object { kind, section }
Affects GET /trademarks/{applicationNumber} ≡ get_trademark when parties are expanded (detail and report), in formerParties[].retiredBy. No other shape changes.
before: "A07T" now: { "kind": "section", "section": "A07T" }
before: "authority_withdrew" now: { "kind": "authority_withdrew", "section": null }
The string mixed two kinds of value, an INAPI section code and a neutral code of our own, and you had to know that authority_withdrew is not a section. The kind now goes in kind and the INAPI code, verbatim and from an open vocabulary, in section; a code not on any list may arrive, and you can interpret it with GET /sections/{code} ≡ get_section. Values do not change: each retired party publishes the same section or the same authority_withdrew as before, only wrapped.
What to do: wherever you read retiredBy as a string, read retiredBy.kind and, if it is "section", retiredBy.section. If you validate the schema strictly, update it to the object.
2026-09-30 · [FIX] Without permission the 403 (and without a key the 401) comes before the validation 400, over REST and MCP
A key without the required scope that sent invalid input received 400 validation_error instead of 403 forbidden, and a request without a key received 400 before 401: the validation error described the input shape to someone not authorised to use it. It happened on every scoped route and tool.
The order is now the same over both transports: 401 → 403 → 400 → burst limit/quota. Invalid input still does not consume quota. Scopes required by an expansion (expand=citations, expand=norms, expand=legalTexts ⇒ legal:read) are decided on the query as received. The 400 for a body that is not readable JSON and the 413 for size still come first. Valid input with permission is unchanged.
What to do: nothing, unless you treat a 400 as a sign that the key and permission are valid.
2026-09-30 · [FIX] MCP: five error codes now carry the status of their REST twin (previously 400)
Over MCP, the envelope of these errors said status: 400 / title: "Request validation failed". status and title are now those of REST:
code |
tools | now |
|---|---|---|
query_timeout (tool's overall limit) |
all | 504 |
watchlist_too_large |
create_watchlist, add_watchlist_item |
422 |
unsafe_callback |
create_watchlist, update_watchlist |
422 |
idempotency_conflict |
create_watchlist |
409 |
coverage_not_declared (previously service_unavailable) |
list_holidays |
503 |
In list_holidays the code also changes, to the same one as GET /holidays.
What to do: if you branch on status over MCP, these cases no longer arrive as 400. If you compare the code of list_holidays, replace service_unavailable with coverage_not_declared.
2026-09-30 · [BEH] Watch store unavailable: 503 service_unavailable over REST and MCP
When the watch store is unavailable, the /watchlists* routes returned 500 internal_error (which hides that a retry may work) and their MCP tools returned service_unavailable with status: 400 (which blames the client). Both now return 503 service_unavailable, with the same neutral detail (The watch store is temporarily unavailable.). It is retryable and the body shape does not change.
What to do: treat a 503 from /watchlists* as transient and retry with backoff.
2026-09-30 · [BEH] REST: a server resource shortage returns 503 service_unavailable (previously 500 internal_error)
A transient failure caused by a storage resource shortage (disk, memory) returned 500 internal_error over REST and 503 service_unavailable over MCP: the same failure was retryable on one side and not on the other. REST now returns the same as MCP: 503, code: service_unavailable, with a neutral detail (The corpus service is temporarily unavailable. Please retry shortly.). Any other internal error is still 500 internal_error. The OpenAPI now documents 503 on every operation.
What to do: treat a 503 as transient (retry with backoff), like a 504.
2026-09-29 · [BEH] MCP: seven catalog and system tools now consume quota and burst limit, like their REST twins
list_legal_norms, get_legal_norm, list_sections, get_section, explain_section_code, get_system_coverage and get_vigilance_horizon did not consume monthly quota or the per-key burst limit, while their REST twins (GET /legal-norms, GET /legal-norms/{id}, GET /sections, GET /sections/{code}, GET /system/coverage, GET /system/vigilance-horizon) did. They now consume one unit per call, as in REST. The scope they require (brands:read for catalogs, insights:read for system) and the response do not change.
In addition, explain_section_code rejects with 400 validation_error a code that is not shaped like a section code (^[A-Z0-9_]{1,10}$ after upper-casing); real codes (024, M11, 664B, A07T…) remain valid.
What to do: nothing, unless you relied on these MCP calls not consuming quota.
2026-09-29 · [BEH] MCP: JSON-RPC batches (an array in the body) are rejected with 400 / -32600
The MCP HTTP transport accepted an array of JSON-RPC messages and ran them all for a single burst unit. A body that is an array (including an empty or one-element array) is now answered with HTTP 400 and {"jsonrpc":"2.0","error":{"code":-32600,"message":"JSON-RPC batch requests are not supported: …"},"id":null}, before authenticating, charging or running anything. The MCP specification removed batches in 2025-06-18, and the official SDK clients do not send them. REST does not change.
What to do: nothing if you use an official MCP client. If a custom client groups messages in an array, send them one per request.
2026-09-29 · [BEH] The search text q (and similarTo) accepts at most 256 characters
q had no maximum length. Now q and similarTo in GET /trademarks, POST /trademarks/search, GET /insights/trademarks, POST /insights/trademarks/search ≡ search_trademarks / get_trademark_insights, and q in GET /parties ≡ list_parties, accept up to 256 characters. More is a 400 validation_error (or the same input error over MCP) with errors[].code = "too_big" on the field. Reports (text, up to 200) and the document base (q, up to 200) already had their limits and do not change.
What to do: if you send text pasted by a user, trim or validate it to 256 characters before calling.
2026-09-29 · [BEH] From 2027-01-01, the Nice class catalog serves the headings of the 13-2027 edition
Until 2026-12-31 (Chile time) nothing changes. From 2027-01-01, list_nice_classes ≡ GET /nice-classes, the resource iris://catalogs/nice-classes and the heading values of recommend_nice_classes ≡ POST /nice-classes/recommend carry the headings of 13-2027, the edition in force from that date.
- 7 headings change: classes 6 and 9 in
es-CLanden, and classes 21, 29 and 31 ines-CLonly. The rest are identical.es-CLandenare verbatim copies from WIPO, without the final period. pt-BRis still Tarno's own, unofficial translation; in 13-2027 it changes in classes 6, 9, 29 and 31.- The recommender, with a reference date from 2027-01-01, now returns
heading["pt-BR"]filled in (previouslynull).
What to do: nothing required. If you cache class titles, invalidate them on 2027-01-01. Read the edition from edition in the recommender response; do not assume it.
2026-09-29 · [BEH] Trademarks no longer lose their category when updated from INAPI's record page (category, protectionScope, get_class_stats)
Thousands of trademarks filed in 2025–2026 came out without a category (13,526 as of 2026-09-29): an update from INAPI's record page, which does not report a subtype, erased the subtype, category and protection scope, and INAPI's spreadsheet restored them the next day.
- A source that does not report subtype, category or scope now does not erase them; if it does report them, it wins.
- Nothing is guessed: a trademark whose source gives no category still comes out without one, and
get_class_statsstill counts it inexcludedUnknownCategory. About 1,100 trademarks of 2026 are like this because INAPI publishes them without type or subtype. - Affected trademarks are fixed automatically with the next spreadsheet update for each year;
excludedUnknownCategoryshould drop within days.
What to do: nothing. If you display excludedUnknownCategory, expect it to drop.
2026-09-29 · [BEH] get_deadlines gives the appeal deadline from the last 024 when later resolutions exist, with a caveat
When a trademark's most recent resolution cannot serve as anchor (M11 —e.g. "dismissed as out of time"—, 664B, 662B, 484, or an M6 without a 024 on the same day) and there is an earlier 024 rejection with no 444 partial acceptance in between, get_deadlines ≡ GET /trademarks/{applicationNumber}/deadlines no longer omits the date (latest_resolution_not_anchoring) and computes it from that last 024, with caveat: "later_resolution_on_record" and the later resolutions in laterResolutions[]. In the corpus as of 2026-09-29, 786 trademarks move from omitted[] to deadlines[].
Unchanged: a 444 as the most recent (partial_acceptance_not_published), cases with no earlier 024 or with a 444 in between, and Madrid, anchor on a non-business day or outside the calendar, which remain omitted. The appeal_close milestone in milestones[] is not published in this case.
What to do: if these trademarks only arrived in omitted[], they now arrive in deadlines[]: read caveat before alerting with that date. It is never a verdict on whether the deadline applies.
2026-09-29 · [ADD] get_deadlines: caveat and laterResolutions on every deadline
Each element of deadlines[] in GET /trademarks/{applicationNumber}/deadlines ≡ get_deadlines gains two always-present fields:
caveat:nullor"later_resolution_on_record"(closed vocabulary).laterResolutions[]: resolutions dated strictly after the anchor, oldest to newest, withannotationId,sectionCode,sectionName(INAPI's name from the catalog, ornull) anddate. It is[]exactly whencaveatisnull.
REST and MCP return the same.
What to do: if you validate TrademarkDeadline strictly, accept both fields.
2026-09-29 · [ADD] Deadline to appeal a rejection: appeal_close milestone and GET /trademarks/{applicationNumber}/deadlines ≡ get_deadlines
The deadline to appeal a rejection (024) before the Industrial Property Court is now published: the 15th business day strictly after its notification in the Estado Diario (excluding Saturdays, Sundays and holidays). Legal basis, in the field: Law 19.039 art. 17 bis B, Law 19.039 art. 11 and Law 19.880 art. 25.
appeal_closemilestone inmilestones[](detail, collection, batch, party, reports):basis: "derived",anchor: {kind: "rejection_notice", date},rule: "15_business_days_after_notification"and the three norms. Not filterable:milestoneKind=appeal_closegives400 validation_error.- New operation (
brands:read):deadlines[]with the auditable anchor (annotationId,sectionCode: "024"),appealObserved(the first laterREC03appeal, a corpus fact),businessDaysLeft/elapsedas of today in Chile;omitted[]with what is not published and why (latest_resolution_not_anchoring,partial_acceptance_not_published,madrid_designation,anchor_on_non_business_day,outside_holiday_calendar);calendar;meta.dataAsOf. Unknown trademark:404 not_found. - Only a
024that is the most recent resolution anchors. Omitted with a reason, never estimated; no safety margin.
What to do: if you validate Milestone.kind strictly, accept appeal_close. Cite the articles and treat omitted[] as "we don't know", not "there is no deadline".
2026-09-29 · [FIX] MCP: a resource that does not exist returns status: 404 (previously 400), as in REST
Over MCP, the error for a resource that does not exist (or, for watchlists, belongs to another organisation) changes from status: 400 / title: "Request validation failed" to status: 404 / title: "Not found", with the detail of the twin REST route (e.g. trademark 123 not found). The body matches REST field by field, except requestId, empty over MCP.
Affects 18 tools: get_annotation, list_trademark_annotations, get_legal_norm, get_section, get_party, list_party_trademarks, list_party_relations, get_portfolio_report, get_opposition_brief, get_opposition_response_brief, get_trademark, get_watchlist, delete_watchlist, list_watchlist_hits, list_watchlist_deliveries, add_watchlist_item, list_watchlist_items and remove_watchlist_item. In update_watchlist only the detail changes. The code is still not_found; REST does not change.
What to do: if you branch on code, nothing. If you branch on status, read 404 as "does not exist (or is not yours)". If you compared the detail text, compare the code instead.
2026-09-29 · [ADD] Declared future operation: POST /logos/similar ≡ find_similar_logos returns 503 image_similarity_not_available
Searching for trademarks similar to a logo is now declared in the contract (input and future output shape), with its own scope logo:similar, but it is not available yet: today it always returns 503 image_similarity_not_available. In the OpenAPI it carries x-availability: planned.
- Input:
image {mediaType, dataBase64}(the same definition asanalyze_logo), optionalniceClasses(1..45) and optionallimit(1..50). - Future output:
results[]withapplicationNumber,denomination,niceClasses,signType,imageUrl,score,signals {visual},matchedBy ["visual"], andmetawithdataAsOf,model,coverage,disclaimerandimage. Not emitted today. - Today: validates the image exactly like
analyze_logo(same400 invalid_image) and returns the503, the same over REST and MCP. The image is not stored. - Without
logo:similar,403. No key holds that scope yet.
Availability will be announced in a separate entry.
What to do: nothing if you don't use it. If you code against it now, treat 503 image_similarity_not_available as "not available yet" and do not retry it; never present it as "no similar trademarks".
2026-09-29 · [ADD] search_annotations by trademark list and by detection time: applicationNumbers, detectedAfter/detectedUntil, detectedAt and meta.detectedThrough
POST /annotations/search ≡ search_annotations lets you watch a list of trademarks and ask "which new acts are there since last time" in one call.
applicationNumbers: up to 100 trademarks per call (duplicates count once). More than 100 ⇒422 batch_too_large; more than 1,000 ⇒400. Not combinable withapplicationNumber,partyIdortaxId.detectedAfter(exclusive) anddetectedUntil(inclusive): ISO 8601 instants with offset about when the corpus learned of it, not the act date. They requireapplicationNumberorapplicationNumbers.detectedAton everyAnnotation: when the corpus learned of the row, not when INAPI published it. Data from the initial load carries its load date. Withdedupe(the default), a fact counts from its first row.meta.detectedThrough: safe horizon, present only with these parameters; nothing detected later is served.
Pattern: request with detectedAfter = the previous detectedThrough; page with cursor and the same detectedUntil = the first page's detectedThrough; store it for next time.
What to do: accept detectedAt if you validate Annotation strictly. To watch lists, use applicationNumbers in batches of 100 with this pattern.
2026-09-29 · [ADD] Annotations quote the term INAPI prints: inapiTermLabel
Every Annotation (POST /annotations/search ≡ search_annotations, GET /trademarks/{n}/annotations, GET /annotations/{id} ≡ get_annotation and the daily bulletin, get_daily_bulletin) gains the always-present field inapiTermLabel: { "basis": "inapi_label", "text": "30 días", "code": "661B", "quotedFrom": "actName" } or null. The annotations inside a trademark's detail or report do not gain it.
It is the text in parentheses of the name INAPI publishes for the act ("Fondo - Res. Observaciones Escrito (30 días)" ⇒ "30 días"), quoted verbatim, only for 14 codes: 661, 661B, 661D, 738, 813, 813A, 813B, 813D, 925, 926, 931, 931B, 931D, 964. It is quoted from actName when there is an act; otherwise from sectionName. It is not a date or a computed deadline, and never enters milestones[].
What to do: accept the field if you validate Annotation strictly. Use it as a quoted fact ("INAPI states 30 days"), never as "due on…".
2026-09-29 · [BEH] analyze_logo ≡ POST /logos/analyze: description is written in the style of INAPI's register label descriptions
The shape does not change; the wording of description.es does, matched to the style measured across about 100,000 label descriptions in the register:
- opens with "Etiqueta consistente en…" ("La marca tridimensional consiste en…" for a three-dimensional shape; a short sentence with "Denominación "…"" for a word-only sign);
- places elements relative to one another, quotes the literal text and ties each colour to its element using names common in Chile ("naranjo", "celeste", "café"), without Pantone/RGB/hex codes, and ends with the background;
- one to three sentences, about 150–300 characters;
- does not write "se reivindican los colores" or phrases about the scope of protection.
Everything else is unchanged: no opinion on registrability, Vienna codes from the catalog, generatedBy and disclaimer.
What to do: nothing required. Keep showing the disclaimer (the description is generated by a model, not by INAPI). Review any text that assumed a different wording (e.g. starting with "Logo…").
2026-09-29 · [BEH] Similarity search: the distinctive channel weighs 2 (previously 1.5) and does not run with a core shorter than 3 characters
Affects POST /trademarks/search with similarTo ≡ search_trademarks, GET /reports/clearance ≡ get_clearance_report, GET /reports/conflict-landscape ≡ get_conflict_landscape and GET /reports/registration-signals ≡ get_registration_signals.
- Ordering changes: the weight of the
distinctivechannel goes from 1.5 to 2, and with it which trademarks enter thedistinctivegroup. Example: "ALTODESEMPEÑO CONSULTORES" / 35 ⇒ "MAS Desempeño" rises from position 77 to 32. - The channel does not run if the distinctive core has fewer than 3 letters or digits (e.g. "AB CONSULTORES" ⇒ "ab"): there is no
distinctivegroup, no trademark carriesdistinctiveinmatchedByandgenericTerms/distinctiveCore/decompoundedare not sent. - No shape changes; only the description of
distinctiveCorein the OpenAPI is updated.
What to do: nothing required. If you store similarity-search positions, expect a different order for texts with generic terms; the distinctive group may be absent for short cores.
2026-09-29 · [FIX] The remaining 400 validation_error responses also carry errors[], one entry per field
400 validation_error responses arising from a cross-field rule (not from a single field's schema) carried only text in detail, without errors[]. They all now carry one entry per field, { "field", "code", "detail" }, with code from a closed vocabulary:
errors[].code |
when |
|---|---|
not_combinable |
the field clashes with another one present; detail says which |
custom |
another contract rule (field not accepted here, field requiring another, text with nothing searchable) |
too_big / too_small |
size or range out of bounds |
unrecognized_keys |
the retired offset parameter |
Covers trademark search (GET /trademarks, POST /trademarks/search ≡ search_trademarks), GET /trademarks/renewal-calendar, GET /insights/trademarks, the /reports/* reports, the document base and statutes, and GET /holidays. The 400 status, the top-level code and the detail do not change. REST and MCP publish the same body.
What to do: nothing required. To know which parameter failed, read errors[].field instead of parsing detail.
2026-09-29 · [BEH] Arts. 6 and 7 of the Law 19.039 Regulation cited before 2022-05-09 are read against DS 82
GET /annotations/{id}?expand=citations ≡ get_annotation with citations. Since late 2017 INAPI cited "art. 6 letter a/b/c/d" and "art. 7 letter b/c" of the Regulation using the numbering of DS 82, which only took effect on 2022-05-09. Those citations came back with text: null and textStatus: "letter_not_found". Now, when the annotation carries the letter's characteristic wording, they are read against DS 82 on its first day: textStatus: "found", scope: "letter", text.asOf: "2022-05-09" (later than the act) and documentCode: "bcn-1175783".
text.chosenBy gains the value later_regulation_template. Art. 7 a) and d), and citations without that wording, do not change. 15,353 citations in 13,879 annotations are affected.
What to do: if you validate text.chosenBy against a closed list, add later_regulation_template. When showing the text, warn that it was not yet in force on the date of the act.
2026-09-29 · [BEH] Art. 9 of the Law 19.039 Regulation that copies the trademark template is read against DS 236, and text.chosenBy says so
GET /annotations/{id}?expand=citations ≡ get_annotation with citations. On 2022-05-09 DS 82 replaced DS 236 and its art. 9 became about patents. INAPI kept citing the DS 236 trademark template, and the rule "the date of the act chooses the text" served the patent article (letter_not_split).
textgains the fieldchosenBy, always present:act_date(general rule) orformer_regulation_template.- An art. 9 citation after 2022-05-08 that carries the trademark-template wording is read against DS 236:
textStatus: "found",text.asOf: "2022-05-08",documentCode: "bcn-244450". 254 citations change.
What to do: if you validate the shape of text strictly, accept the chosenBy field.
2026-09-29 · [BEH] The opposition freshness note explains how long an opposition act takes to arrive
The text of freshnessNote changes; no schema changes. It appears in meta of GET /trademarks/{applicationNumber}/annotations ≡ list_trademark_annotations and of POST /annotations/search ≡ search_annotations when the opposition family is requested, and as a prefix of the notes of get_class_stats and get_registration_signals.
The note now says that the first Estado Diario C1 arrives between 55 and 375 days after publication, so the absence of opposition acts during the window, and months later, does not prove there is no opposition. The window (publication + 30 business days, art. 5 Law 19.039) remains the only derived deadline. The contract version (info.version in the OpenAPI and serverInfo.version in MCP) moves to 2026-09-29.
What to do: nothing if you show the note as is. If you compare it with a literal, update it or only check that it exists.
2026-09-29 · [BEH] get_filing_context (opposition): a trademark without a comparable denomination omits the comparison, not the file
GET /reports/filing-context ≡ get_filing_context with type=opposition and ownApplicationNumber: if one of the two trademarks had no comparable denomination (figurative or symbols only), the whole file answered 422 no_comparable_denomination. Now it arrives complete, with comparison: null and one comparison.omitted fact in facts per affected trademark (value no_comparable_denomination, source.kind: "trademark_comparison", source.ref = its application number). A non-existent own trademark still answers 404.
What to do: if you assumed comparison is never null when passing ownApplicationNumber, check comparison.omitted in facts.
2026-09-29 · [BEH] analyze_logo: an image the provider cannot process is 422 logo_image_rejected, not a 503
POST /logos/analyze ≡ analyze_logo. Local validation only reads the image header: a truncated file, or garbage after a valid header, passes it and the provider rejects it. Until now that answered 503 logo_analysis_unavailable ("retry later"), and each retry of the same image spent quota. It now answers 422 logo_image_rejected on both transports, not retryable. 503 logo_analysis_unavailable is kept for real unavailability. The 422 counts against the daily and the monthly quota.
What to do: treat 422 logo_image_rejected as an image error (ask the user for another one) and do not retry it.
2026-09-29 · [BEH] The generic-term detector has caps, and frequency.notEvaluated declares what it did not examine
Affects the distinctive channel of POST /trademarks/search ≡ search_trademarks with similarTo, the availability report, POST /trademarks/compare ≡ compare_trademarks and GET /reports/filing-context ≡ get_filing_context. Without changing the input:
- a word longer than 30 letters is not examined;
- only the first 12 distinct words of the text are examined;
- there is a maximum number of probes per detection.
What is left unexamined goes to frequency.notEvaluated and, with classes, frequency.state is partial. In similarity search, that word simply does not appear in decompounded or genericTerms and stays in distinctiveCore.
What to do: nothing. If you show notEvaluated, its description explains the reasons.
2026-09-29 · [ADD] Logo analysis with a model: POST /logos/analyze ≡ analyze_logo
Requires its own scope, logo:analyze. It is the only operation in the contract that uses a language model (Anthropic), and the response declares it in generatedBy.
- Input:
image {mediaType, dataBase64}: PNG, JPEG, GIF or WebP, up to 700 KB and 2,000 px per side. Never a URL (400). - Response:
disclaimer,description,elements,suggestedSignType,viennaCodes[](only codes that exist in VCL10, with the WIPO label in English and French;esalwaysnull),discardedCodes,viennaCatalogandimage(withsha256; the image is not stored). - Own quota: 20 per key per Chilean day, 200 per day in total ⇒
429 logo_quota_exceeded/429 logo_global_quota_exceededwithRetry-After. - Errors:
400 invalid_image(withreason),422 logo_analysis_refused,502 logo_analysis_invalid_output,503 logo_analysis_not_configured,503 logo_analysis_unavailable,503 logo_quota_unavailable.
No key gets it by default. It gives no opinion on registrability.
What to do: to use it, grant logo:analyze to the key by writing the full list of its scopes. Always show the disclaimer and the WIPO attribution; treat 503 logo_analysis_not_configured as "not available yet" and handle the daily 429 separately from the monthly one.
2026-09-29 · [ADD] Fact file for a filing: GET /reports/filing-context ≡ get_filing_context
A report with its own scope, reports:filing-context:read (also granted by reports:read).
- Input:
type(power_of_attorney,opposition,opposition_response,office_action_response,withdrawal),applicationNumber,annotationId(required withoffice_action_response),ownApplicationNumber(oppositiononly) andexpand=legalTexts. - Response:
trademark,facts[]({key, value, source}),milestones,milestonesOmitted[],citedNorms[],normSignals[],comparison,outline(structure without text),dataAsOf,coverageThrough. - The only published deadline is the opposition close (
opposition_close, art. 5 Law 19.039). Deadlines without a verified article go tomilestonesOmittedwithno_legal_basis_in_knowledge_base. expand=legalTextsalso requireslegal:read; without it,403 forbiddenwithout consuming quota.- It does not draft or give verdicts.
third_party_observationsis not a type (400). Non-existent trademark ⇒404.
What to do: to use it, grant reports:filing-context:read (or reports:read) and, if you want the legal texts, legal:read, writing the full list of scopes. Do not show as a deadline anything that is not in milestones.
2026-09-29 · [ADD] Nice class recommendation: POST /nice-classes/recommend ≡ recommend_nice_classes
Scope brands:read. Input: description (3–500 characters) and limit (1–10, default 5); stop words only ⇒ 400 validation_error.
Response: edition (the one in force in Chile: 13-2026 until 2026-12-31, 13-2027 from 2027-01-01), editionSelection, listVersion, candidates[] (class, heading in three languages, signals, ranks, up to 5 officialTerms[] and 3 corpusExamples[]), method, meta {channels, pendingEmbeddings}, attribution (WIPO, CC BY 4.0) and dataAsOf. Each path declares its state in meta.channels: ran, degraded, not_configured, pending_embeddings or not_loaded.
What to do: nothing if you do not use it. If you show it, cite attribution and the edition, and treat not_loaded/pending_embeddings as "not yet", not as an error. It is a signal, not an official classification.
2026-09-29 · [ADD] Registration signal index: GET /reports/registration-signals ≡ get_registration_signals
A report with its own scope, reports:registration-signals:read (also granted by reports:read; reports:clearance:read is not enough).
- Input:
text(1–200) andniceClass(1 to 3 classes; 4 ⇒400). - Response:
perClass[]withsignals[](S1…S6, each with value, source and evidence),signalIndex(0–100 ornull) andweightsApplied; plusformula,weights(provisional),indexLabel("not a verdict", three languages),method, the generic-term discount,dataAsOfandcoverageThrough. - A null signal is left out and its weight is redistributed. Before the first computation of class statistics, S3 and S4 come back
nullwithclass_stats_pending.
What to do: to use it, grant reports:registration-signals:read (or reports:read). Show the signals with their source and always indexLabel next to the index; do not turn it into "low/medium/high".
2026-09-29 · [ADD] Conflict landscape: GET /reports/conflict-landscape ≡ get_conflict_landscape
A report with its own scope, reports:conflict-landscape:read (also granted by reports:read). Input: text (1–200) and niceClass (1 to 10).
It takes the 60 most similar trademarks from the same engine as get_clearance_report, keeps the live ones and groups them by the parties with a current relation. Response: representatives[] and holders[] (up to 20), each with party, similarTrademarks, evidence[] (≤ 5), evidenceTruncated and totalTrademarksInClass[]; plus partiesTotal, method, the generic-term discount, dataAsOf and coverageThrough. Without the scope, 403 forbidden without consuming quota.
What to do: to use it, grant reports:conflict-landscape:read (or reports:read). Always show totalTrademarksInClass next to similarTrademarks: these are signals, not a conclusion about a party.
2026-09-29 · [ADD] Class statistics: GET /nice-classes/{niceClass}/stats ≡ get_class_stats
Scope brands:read. Input: the class (1–45) and filedFromYear/filedToYear (default: the last five full years). Inverted range ⇒ 400 invalid_range.
Response: period, population, outcome groups with grantShare, refusedShare, withdrawnShare and annulledShare (same definition as get_party_outcomes), timeToRegistration (approximate percentiles, ±15 days), actsShare, coverage {notes[]}, definition and dataAsOf. Figures are recomputed on each sync; before the first computation, 503 class_stats_pending.
What to do: nothing if you do not use it. If you show it, cite population.filed and the period, say the times are approximate and read coverage.notes. Treat 503 class_stats_pending as "not computed yet".
2026-09-29 · [BEH] Similarity search and availability discount generic terms: distinctive channel
POST /trademarks/search ≡ search_trademarks with similarTo and GET /reports/clearance ≡ get_clearance_report gain a fourth channel, distinctive: it searches for the distinctive core of the text, without the generic terms of its classes and with run-together words split.
matchedBy[]andsimilar[].channelgain the valuedistinctive.- New optional fields, only when the channel ran:
genericTerms[],distinctiveCore,decompounded[]. If there is nothing to discount, the response is the same as before. signals,score,similarToTrademarkand watch do not change.
What to do: if you validate matchedBy or similar[].channel against a closed list, accept distinctive. A term being generic is not a verdict.
2026-09-29 · [BEH] The semantic channel breaks ties by application number: the same similarTo gives the same order
POST /trademarks/search ≡ search_trademarks (with q, similarTo or similarToTrademark), the reports that look for similar trademarks, and watch. Two trademarks with the same vector (e.g. unnamed sound marks) tied in the semantic channel in a non-deterministic order, which could change their score and the order of items between two identical calls. Ties are now broken by ascending applicationNumber: among trademarks with the same vector, the lower number scores slightly higher, always the same one. The order of non-tied trademarks does not change. The response shape does not change.
What to do: nothing.
2026-09-29 · [FIX] The batch 400 for unsupported parameters carries errors[], one per field
GET /trademarks, POST /trademarks/search ≡ search_trademarks. applicationNumber combined with what the batch does not support (cursor, limit, sort, minScore, q, denomination, facets, similarTo, similarToTrademark, allClasses) answered 400 validation_error without errors[]. It now carries one entry per unsupported field, in a stable order:
{ "field": "cursor", "code": "not_combinable", "detail": "…" }
The status, the top-level code and detail do not change. REST and MCP return the same body.
What to do: nothing required. If you currently parse detail to find the extra field, you can read errors[].field.
2026-09-29 · [DATA] Norm catalogue: 39 norms, 38 with a title
/legal-norms, /legal-norms/{id} ≡ list_legal_norms / get_legal_norm, and the citedNorm filter. After the catalogue clean-up: norm id 48 (a non-existent "reglamento 19.880"; its citations move to the Law 19.039 Regulation) is removed, as are 15 norms that were typos, whose citations are re-pointed to Law 19.039, its Regulation or Law 19.880. Norm id 112 (Law 20.000 Regulation) gets its official title, and laws 1.552, 19.496 and 19.996 are added. Only law 19.037 (id 59) remains without a title, on purpose. The shape does not change.
What to do: a stored id of a removed norm answers 404; if you list the norms, four more appear with a title.
2026-09-29 · [DATA] act stops being null in 103 C7/C8 annotations
Annotation, REST ≡ MCP. 103 C7/C8 annotations whose heading had been stored truncated recover the full heading from the archived bulletin, which identifies their act: they move from act: null / actStatus: "unidentified" to 738 (47), 741 (33), 919 (9), 964 (8) and 925 (6), with their actName/actLabel and actStatus: "identified". No other field changes. Other annotations stay unidentified because the bulletin does not allow identifying them.
What to do: nothing; a count of annotations without an act drops by 103.
2026-09-29 · [DATA] act stops being null in 592 family-A annotations
Annotation, REST ≡ MCP. The 592 family-A annotations whose observation was completed (entry below) move from act: null to their INAPI code, read from the subsection heading of the archived bulletin, with its name and label; their actStatus becomes identified. No other field changes.
What to do: nothing; a count of annotations without an act drops by 592.
2026-09-29 · [DATA] Citations: the Madrid Protocol and the Civil Code no longer count as Law 19.039; the CPC moves to Law 1.552
citedNorm, /legal-norms ≡ list_legal_norms / get_legal_norm, expand=citations and the Guidelines citations. Citations of arts. 4 and 5 of the Madrid Protocol and 16 and 17 of its Common Regulations were recorded as articles of Law 19.039 and its Regulation; they stop being citations, as do those of the Civil Code (≈44,100 citations in total). Those of the Code of Civil Procedure move to Law 1.552 (≈11,200). In the Guidelines, ≈105 citations move to 1.552 and "artículo 6 ter del CP" (Paris Convention) stops being a citation. The shape does not change.
What to do: redo any count of citedNorm=19039 with article 4 or 5, of the Regulation with 16 or 17, or of citations without a law; citedNorm=1552 returns the CPC ones.
2026-09-29 · [ADD] Three new laws in the catalogue: 1.552 (CPC), 19.496 and 19.996
/legal-norms ≡ list_legal_norms / get_legal_norm, get_legal_norm_article. They arrive with their official BCN title: Law 1.552 "Código de Procedimiento Civil", Law 19.496 (consumer rights protection) and Law 19.996 (amends Law 19.039). Their text is loaded with all its versions in the weekly BCN update. CPC articles answer xml_missing, because BCN does not serve its structured text. In /legal-norms only 1.552 has articles.
What to do: nothing; if you list the norms, three more appear.
2026-09-29 · [BEH] Cited transitional articles: article = '3 transitorio' and its text is the transitional one
GET /annotations/{id}?expand=citations ≡ get_annotation with citations, and the citations of /guidance/*. "Artículo 3° transitorio de la Ley 19.996" was published as permanent article 3, with its text. Now:
articlecarries the suffix ('3 transitorio'; "octavo transitorio" ⇒'8 transitorio'). Without the suffix it is still the permanent article.textis the transitional article's (found), orarticle_not_found, orambiguous_articleif two match. Never the permanent one.- A transitional article without a written law (e.g. company bylaws) is no longer published as a citation.
15 citations are added and 19 that were transitional are removed. The shape does not change.
What to do: if you parse article as a number, accept the transitorio suffix.
2026-09-29 · [DATA] Family-A annotations: full observation where it used to be truncated
Annotation, REST ≡ MCP. 592 family-A Estado Diario annotations (A1 391 · A4 156 · A3 45; 2024-09-30 → 2026-08-07) had their observation truncated at the bulletin's page break. They now carry the full text: the new one starts literally with the previous one and continues (from ~95 to ~1,831 characters on average). The id, parties[] and the other fields do not change; new citations found in the added text may appear. The shape does not change.
What to do: nothing; if you store observation, refresh it.
2026-09-29 · [BREAK] extractionStatus gains the value ocr, and every search_guidance hit carries it
An INAPI document published only as an image (for example, chapter 21 of the Guidelines, a diagram) could be image_only, with no text. After machine reading it becomes ocr: it has one section per page (path ocr.p1, kind block) that is read with get_guidance_section and found by search_guidance, but its text is not literal: some words may be misread and diagram labels come out in page order.
Two changes in /guidance/* ≡ list_guidance_documents, get_guidance_document, get_guidance_section, search_guidance:
extractionStatusacceptsocrbesidespending,extracted,partial,image_onlyandfailed. It breaks only clients that validate the closed list (an exhaustiveswitchor a strict enum).- Every hit of
POST /guidance/search≡search_guidancecarriesversion.extractionStatus(new field, always present). Itsexcerptis literal except when that value isocr.
What to do: accept ocr and, when it appears, present the text as "machine-read from an image" and link to the PDF (sourceUrl) for quoting. No existing field changes name or type.
2026-09-29 · [DATA] partial now means only "incomplete tree": 18 Guidelines chapters move to extracted
Until now a chapter was also partial when a section heading was worded differently in the table of contents and in the body. Now partial means only an incomplete tree: a table-of-contents entry with no body, or a body heading the table of contents does not list. In addition, a heading printed at body-text size is now recognised and the page footer no longer leaks into the text.
Across the 50 chapters: from 18 extracted / 31 partial to 35 extracted / 14 partial (chapter 11 moves down to partial). Paths and text are identical in 47 chapters; chapter 10 gains II and its § 3 and § 4 move from I.B.§3/I.B.§4 to II.§3/II.§4; chapters 9 and 32 lose 9 boxes (…sidebarN) that were only the page footer, and the following ones are renumbered.
What to do: if you stored references to those sections, resolve them again by (code, version, path).
2026-09-28 · [ADD] Compare two trademarks path by path: POST /trademarks/compare ≡ compare_trademarks
Scope brands:read. Input: a and b, each exactly one of {text, niceClass?} or {applicationNumber}. Invalid combination ⇒ 400 validation_error; non-existent number ⇒ 404; trademark without a comparable denomination ⇒ 422 no_comparable_denomination.
Response: per side, text, class, phonetic key, genericTerms[], decompounded[] and distinctiveCore; signals (lexical, edit distance, phonetic, semantic, shared tokens), distinctiveSignals, classes, goodsAndServices (trademark vs trademark only), method, meta and dataAsOf. No overall score, no level, no "risk". semantic: null means not evaluated, never 0.
What to do: show the paths separately with their method and, if you summarise, cite genericTerms. Do not combine the signals into a percentage.
2026-09-28 · [ADD] search_trademarks: partyRole restricts taxId to one role
GET /trademarks, POST /trademarks/search and search_trademarks accept partyRole (holder | representative) together with taxId: the RUT matches only trademarks where the party has that role, in every engine. Without partyRole nothing changes. partyRole without taxId ⇒ 400 validation_error, identical in REST and MCP. Same vocabulary as partyRole in search_annotations.
What to do: nothing if you do not use it. For "the trademarks this RUT represents", add partyRole=representative.
2026-09-28 · [ADD] search_annotations: filters by party and by the trademark's current status
POST /annotations/search ≡ search_annotations accepts five optional fields:
partyIdortaxId(at most one): only annotations of that party's trademarks. Non-existentpartyId⇒404;taxIdnot shaped like a RUT ⇒400 invalid_rut.partyRole(holder|representative) andrelation(currentby default |former|all): only with a party; otherwise400 validation_error.trademarkStatus: only annotations of trademarks whose current published status is in the list (same semantics asstatusin trademark search).
Without these fields the response does not change.
What to do: nothing if you do not use it. "What happened this week to this RUT's trademarks?" is now { "taxId": "…", "dateFrom": "…" }.
2026-09-28 · [ADD] A party's outcomes: GET /parties/{id}/outcomes ≡ get_party_outcomes
Scope brands:read. Input: role (representative by default, or holder), relation (all by default), niceClass, filedFrom, filedTo.
Response: party, population, groups[] (registered_ever, refused, withdrawn_or_abandoned, annulled, pending, each with its list of statuses), unclassified, decided, pending, grantShare, refusedShare, withdrawnShare, annulledShare (over decided ones; null, never 0, if there are none), definition in three languages and dataAsOf. Non-existent party ⇒ 404; filedFrom > filedTo ⇒ 400 invalid_range. It is not a "success rate": it is a descriptive signal.
What to do: if you show it, accompany the ratio with population.trademarks and the definition.
2026-09-28 · [ADD] Official INAPI fees: POST /fee-schedule/calculate ≡ calculate_fees
Scope brands:read (the full table, GET /fee-schedule, stays under legal:read).
- Input:
procedure(application_and_registration|renewal), exactly one ofniceClassCount(1–45) orapplicationNumber,on(payment date) andutmMonth. - Response:
lines[]with amounts in UTM and pesos,totalUtm,totalClp,totalOmittedReason(surcharge_doubt|lapsed|null),utm,renewal,excluded,trademark,source,dataAsOf. - During the renewal grace period the total is
null(surcharge_doubt); after it,lapsed. - Errors:
404,422 renewal_not_applicable,422 no_nice_classes.
What to do: if you show costs, show excluded and, with totalOmittedReason, explain why there is no total instead of adding up the lines.
2026-09-28 · [BEH] MCP: a scope rejection comes back with status: 403 (was 400), as in REST
Over MCP, a tool called with a key lacking the required scope changes from status: 400 / title: "Request validation failed" to status: 403 / title: "Forbidden", with the same detail as REST (API key lacks required scope: brands:read). A missing, unknown or revoked key changes from 400 to 401 / "Unauthorized". The code does not change (forbidden / unauthorized). REST does not change.
What to do: if your MCP client branches on code, nothing. If it branches on status, read 403 as "missing scope" and 401 as "invalid key", not as an input error.
2026-09-28 · [ADD] Business-day calculator: GET /business-days ≡ compute_business_days
Scope brands:read. Input: from (YYYY-MM-DD) and exactly one of businessDays (1–250) or preset (opposition only); optional anchorKind. Invalid combination ⇒ 400 validation_error.
Response: anchor, businessDays, date (the Nth business day strictly after from, skipping weekends and holidays), term, legalBasis, calendar, skipped, basis (derived | estimated), estimateReasons, todayChile, elapsed and businessDaysLeft. Without a preset, term.source: "caller" and term.legalBasis: null: a number is not a deadline. With preset=opposition (art. 5 Law 19.039, N = 30) it cites its norms.
What to do: show the date with legalBasis, say "estimate" when basis is estimated, and do not present a count with term.source: "caller" as a legal deadline.
2026-09-28 · [ADD] Norms cited in the trademark record: GET /trademarks/{n}?expand=norms ≡ get_trademark
norms was already an accepted expand value but returned nothing. The trademark record now carries two always-present fields: norms and normsTruncated (both null when not requested). norms[] has one entry per (kind, norm, article, letter, subsection) cited in the trademark's annotations: normKind (ley|reglamento), legalNormId (null when the act cites an article without naming the law; that citation is still included), citedNorm, name, article, letter, subsection, annotationCount, firstCitedOn, lastCitedOn and annotationIds (up to 20, the most recent). Most-cited first; at most 100 entries (normsTruncated: true if there are more). With dedupe=true (the default) annotationCount counts events, not rows.
The statute text is not included here: read it with GET /annotations/{id}?expand=citations using one of the annotationIds. Requesting norms requires legal:read in addition to brands:read; without it, 403 forbidden naming legal:read, before any quota is consumed. Until now the same request returned 200.
What to do: if you request expand=…,norms with a key lacking legal:read, drop norms or request the scope. To show a record's norms, read norms[] instead of walking annotations[].
2026-09-28 · [BEH] Documentary sync-run error no longer includes the full database message — GET /system/sync-runs ≡ list_sync_runs
In runs with source inapi_documents, bcn_statutes and sii_indicators, a database failure appeared in error with the full message, which could carry values from the failing row. It now shows only a standard error code, plus the table, constraint or column names when available. A connection error shows only its code (ECONNREFUSED). Other sources do not change and error is still free text.
What to do: nothing, unless you parse the error text; the overall format (code (reason); …) does not change.
2026-09-28 · [ADD] Explained citations: GET /annotations/{id}?expand=citations ≡ get_annotation with expand
An annotation's detail can include, for each norm the decision cites, the literal text of the article or letter in force on the annotation date and the signals from the Guidelines. The route accepts only expand=citations; it requires legal:read in addition to brands:read (without it, 403 forbidden before any quota is consumed).
The response adds citations[] (up to 50, in the decision's order) and citationsTruncated. Each citation has normKind, legalNormId, citedNorm, name, article, letter, subsection, span; text (asOf, documentCode, validFrom, validTo, articleLabel, scope, body…) or null, with the same validity rule as get_legal_norm_article; textStatus (found, letter_not_split, law_not_identified, no_text_loaded, no_date, out_of_window, article_not_found, ambiguous_article, letter_not_found); and guidance[] (up to 5 INAPI documents citing that article, with their sections). These are signals: no field says which section interprets the citation. In lists and in GET /trademarks/{n}, expand=citations is 400 validation_error. Without expand, the response is identical to before.
What to do: nothing if you do not show norms. If you do: request expand=citations only when opening an annotation, show text.asOf next to the text and textStatus when there is no text.
2026-09-28 · [ADD] Fees, indicators and version changes: get_fee_schedule, list_economic_indicators and list_guidance_changes
Three new documentary-base reads, scope legal:read:
GET /fee-schedule≡get_fee_schedule: the fee table from chapter 13 of the Guidelines, in UTM. InpututmMonth?(YYYY-MM). Returnsitems[]withamountUtm,unit(per_class|fixed),amountClp(computed on read:amountUtm × utm.value;nullwithout UTM),hasDoubt, plusutm {month, value, basis}andsurchargeDoubt, which exposes the open question about the late-renewal surcharge. No table loaded:422 out_of_coverage.GET /economic-indicators≡list_economic_indicators: requiredcode(UTM,UTA,IPC,UF,USD_OBSERVED),from/to,limit(1–366),cursor. Rows{code, date, value, unit, baseNote}andmeta.coverage. Range outside coverage:422;from > to:400.GET /guidance/changes≡list_guidance_changes: version changes of INAPI documents and BCN texts, filtersdocument,publisher,from/to; each change hasfromVersion/toVersion,detectedAt,countsandsections[](up to 100).
What to do: nothing if unused. Otherwise: the key needs legal:read; when showing pesos, also show utm.month, and do not mix IPC values with different baseNote.
2026-09-28 · [ADD] Searching the documentary base: search_guidance ≡ POST /guidance/search and search_statutes ≡ POST /legal-norms/search
Two new searches, scope legal:read. search_guidance searches the sections of the current version of each INAPI document (q, mode, kind, document, limit, cursor); each hit carries document, version, path, heading, paragraphNumber, pages, excerpt and matchedBy. search_statutes searches statute articles in the text in force today; asOf picks a date and allVersions: true searches every validity period (identical text appears once, with all periods in validity[]). norm narrows to one norm.
mode: hybrid (default, text + meaning similarity), text (supports "phrase", or, -word) or semantic. meta.channels states which channel ran; if the semantic channel is unavailable, hybrid returns text results only with semantic: "unavailable" and semantic returns 503 service_unavailable. excerpt is literal, with no scores. Edges: q with no searchable words or only exclusions, or asOf together with allVersions: 400; unknown norm: 404; no text loaded: 422 out_of_coverage.
What to do: nothing if you do not search. Otherwise: request legal:read, show validity on statute hits and warn when the search ran on words only.
2026-09-28 · [ADD] INAPI Guidelines, readable: get_guidance_section, list_guidance_documents and get_guidance_document ≡ /guidance/documents/*
Documents INAPI publishes (Guidelines chapters, circulars, Madrid manual and page, fee publications) can be read through the contract, scope legal:read:
GET /guidance/documents/{code}/section≡get_guidance_section: one literal section with its pages, itsversion(withextractionStatus), itschildren[]and the norms it cites (citations[]). Selected bypathor byparagraph, exactly one.GET /guidance/documents≡list_guidance_documents: the collection, withcurrentVersionand filterskind(guideline,circular,manual,web_page,fee_schedule) andq.GET /guidance/documents/{code}≡get_guidance_document: the document, all itsversions[]and its table of contents (outline[], without bodies).
A section's stable reference is (code, version.seq, path). An image_only version returns 422 out_of_coverage when a section is requested. Both or neither of path and paragraph: 400; repeated paragraph number: 409 ambiguous_selector with candidates; not found: 404. dataAsOf belongs to the documentary source, not to the trademark corpus.
What to do: nothing if you do not show the Guidelines. Otherwise: request legal:read, cite by code + version.seq + path and show extractionStatus when it is not extracted.
2026-09-28 · [ADD] get_legal_norm_article ≡ GET /legal-norms/{id}/articles/{article} and the legal:read scope
New legal:read scope for the documentary base, denied by default: a key with empty scopes, the trial key, reports:read or brands:read do not get it and receive 403 forbidden naming it.
The tool returns the literal text of a statute article. Input: id (from /legal-norms), article (20, 17 bis a), and optional letter, subsection, asOf (default today, Chile time) and transitional. Output: norm, asOf, text (which BCN text was read and its validity), article, articleLabel, headingPath, body, letters[], subsections[], notes[], selection and dataAsOf. Errors: unknown norm or missing article, 404; norm with no text loaded or date out of validity, 422 out_of_coverage; two articles with the same key, 409 ambiguous_selector (new code) with candidates; letter requested on an unsplit article, 200 with the whole article and selection.status: "article_not_split".
What to do: request legal:read for the key that reads the documentary base. When granting it, write the complete scopes list: going from [] to ["legal:read"] would remove every other read.
2026-09-28 · [DATA] Three new documentary sources in GET /system/sync-runs ≡ list_sync_runs
Three new source values may appear: inapi_documents (weekly), bcn_statutes (weekly) and sii_indicators (daily). They store what INAPI, BCN and the SII publish. The trademark corpus dataAsOf does not change: these runs do not count towards it. source was already open text.
What to do: nothing, unless you show sync status: filter by source, since an SII run says nothing about trademark freshness.
2026-09-27 · [DATA] act identifies the two C15N acts (832 and 833)
The two subsection headings of C15N ("Cúmplase en juicios de Nulidad") are now recognised as acts 832 and 833. The 8 annotations that came out with act: null and actStatus: "unidentified" now have act 832/833 and identified. 1,993 annotations remain unidentified. The shape does not change.
What to do: nothing.
2026-09-27 · [DATA] Norms: 19 more titles, 5 misspellings of 19.039 and 5 empty regulations removed
The /legal-norms ≡ list_legal_norms catalogue goes from 63 to 51 norms, 34 with a title. Nineteen laws get their official BCN title (among them 21.045, 21.094, 20.606, 20.285, 17.336). Twelve norms cease to exist and their id returns 404 on /legal-norms/{id} ≡ get_legal_norm: the mistyped variants of 19.039 (1039, 13039, 15039, 18039, 1903), whose 16 citations move to 19.039, and 5 regulations with no citations (18.445, 17.336, 18.455, 21.355, 19.031). Another 10 duplicate citations stop being counted under other laws. The shape does not change.
What to do: a stored id of those 12 norms no longer resolves; citedNorm with 1039, 13039, 15039, 18039 or 1903 no longer returns results (use 19039).
2026-09-26 · [BEH] C13N and C15N gain a name and label
Two nullity-series codes had no name. They now carry the literal Estado Diario heading: C13N "Ejecutorias en juicios de Nulidad" and C15N "Cúmplase en juicios de Nulidad". Their 16 annotations go from sectionName: null to having a name; sectionLabel carries es-CL in both and en/pt-BR in C13N (null in C15N). source.label/label in /sections ≡ list_sections/get_section change too. The catalogue now has 674 codes, 543 named. Catalogues may take up to a day to reflect this because of caching.
What to do: nothing.
2026-09-26 · [BEH] POST /annotations/search ≡ search_annotations with section= no longer returns 504
With the section= filter, annotation search timed out (504 query_timeout) for sections that start late in the corpus (M5, ARENTLT, D5A, M11, C2…). It now responds normally. Shape, order (date ASC NULLS LAST, id ASC) and cursors do not change.
What to do: nothing. If you avoided section= or retried, you can now use it directly.
2026-09-25 · [BEH] GET /parties/{id}/trademarks ≡ list_party_trademarks filters by the party's identity, not its name
With the holder or representative role, the list filtered by a name substring: the portfolio of "ACME" also returned the trademarks of "ACME CHILE". It now filters by the party's identity, so it returns only the trademarks the party holds or represents. The shape does not change; the returned set does. Without a role and with a RUT, behaviour is unchanged.
What to do: nothing. If you relied on receiving trademarks of similar names, they no longer arrive, which is correct.
2026-09-25 · [BREAK] expiresOn removed from the annotation: it carried the filing date, not an expiry
The annotation (GET /trademarks/{n}/annotations, POST /annotations/search, the trademark record, the reports and their MCP twins) published expiresOn, but in almost every row that value was the trademark's filing date, not an expiry. Read as an expiry, it produced false alerts. The field is removed.
The event date stays in date. The trademark's expiry (and its expiry milestone) stays in the trademark summary's expiresOn, unchanged.
What to do: if you read expiresOn on an annotation, it no longer exists: use date for the event date or the trademark's expiresOn for the expiry.
2026-09-25 · [DATA] 18,951 annotations linked to their parties: parties[] picks them up
18,951 bulletin names were linked to their parties: 16,552 resolved to a party and 2,399 flagged as ambiguous or without a candidate. Among them, 103 Estado Diario annotations that had recovered their representative and applicant. Those annotations now carry parties[].
What to do: nothing; re-read parties[] of recent annotations if you cached them empty.
2026-09-25 · [ADD] get_renewal_calendar ≡ GET /trademarks/renewal-calendar: the renewal calendar of a set of trademarks
Input: the trademark filter of GET /trademarks (an applicationNumber batch of up to 100, renewalStatus, holder, niceClass, status, dates…) plus required from/to (YYYY-MM-DD, inclusive, at most 366 days), limit and cursor. No range or more than 366 days: 400 validation_error; from > to: 400 invalid_range; batch of 101: 422 batch_too_large; q, similarTo, similarToTrademark, allClasses, oppositionWindow, facets, citedNorm, article and letter: 400.
Output: one row per (trademark, renewal milestone) in the range, ordered by date and number: {applicationNumber, kind, date, basis, anchor, rule, legalBasis}, with kind renewal_window_open or grace_end; each row is the same milestone the trademark publishes in milestones[]. Exact total. Scope brands:read.
What to do: nothing if unused. Example: GET /trademarks/renewal-calendar?applicationNumber=…&from=2026-01-01&to=2026-12-31.
2026-09-25 · [ADD] groupBy=renewalStatus and the portfolio by renewal status
GET|POST /insights/trademarks ≡ get_trademark_insights accept groupBy=renewalStatus: one {key, label, count} group per status (not_due, window_open, grace, lapsed), with labels in three languages that describe the term, not the legal status. meta.undatedCount (trademarks without an expiry date) also travels with this groupBy, and meta.notApplicableCount (new) counts dated trademarks to which the rule does not apply. The sum of count plus both fields is the filtered total.
GET /reports/portfolio ≡ get_portfolio_report gains byRenewalStatus {groups, undatedCount, notApplicableCount}, consistent with each trademark's renewalStatus.
What to do: nothing if unused. When charting the split, show undatedCount ("no date") and notApplicableCount ("not applicable") separately, never as a status.
2026-09-25 · [ADD] renewalStatus filter and renewal milestoneKind values (renewal_window_open, grace_end)
renewalStatus is a new trademark-filter parameter (GET /trademarks, POST /trademarks/search ≡ search_trademarks, GET|POST /insights/trademarks ≡ get_trademark_insights): a list (OR) of not_due, window_open, grace, lapsed, evaluated on the response date (Chile time). A trademark returned by renewalStatus=grace carries renewalStatus: "grace"; those with null match no value. Any other value is 400 validation_error. It works in the collection, search, batch, similarity, opposition window and insights.
milestoneKind also accepts renewal_window_open and grace_end (previously 400), and matches exactly the date the trademark publishes in milestones[].
What to do: nothing if unused. Example: GET /trademarks?applicationNumber=…&renewalStatus=grace.
2026-09-25 · [BEH] milestoneFrom/milestoneTo without milestoneKind also include renewal milestones
Without milestoneKind, the milestone filter matches trademarks with any filterable kind in the range. There are now four: expiry, opposition_close, renewal_window_open and grace_end. A call without a kind also returns trademarks whose only date in the range is the opening of the renewal window or the end of the grace period (for October 2026: 12,905 trademarks, versus 8,056 before). The portfolio does not change.
What to do: if you filter by milestones expecting only expiries (or expiries and opposition closes), pass milestoneKind explicitly.
2026-09-25 · [DATA] Parties: 1,843 duplicate records merged into 1,480
Some parties were duplicated because a name was stored with line breaks or double spaces (same entity, same country, at most one RUT in the group). In each group one record stays and the others merge into it: 1,480 groups, 1,843 merged records, 524 trademark–party links moved (no trademark loses its holder) and 2 parties[] links in annotations. This affects /parties, /parties/{id}, /parties/{id}/trademarks ≡ list_parties/get_party/list_party_trademarks, the parties in the trademark record and search by party.
Merged records still exist with trademarkCount: 0; their /trademarks is empty. Parties with different RUTs were not touched.
What to do: re-measure anything that depends on per-party counts. If you stored the id of a merged record, its trademarks are now on the other record with the same name.
2026-09-25 · [ADD] renewalStatus and the renewal_window_open / grace_end milestones (art. 24, Law 19.039)
New field renewalStatus (not_due | window_open | grace | lapsed | null, always present) in the trademark record (GET /trademarks/{n} ≡ get_trademark) and in every trademark of the collection, search, batch, similarity, window, list_party_trademarks and the clearance and portfolio reports. It is anchored on expiresOn and evaluated on the response date, Chile time. Current rule (expiries from 2022-05-09): not_due before expiry − 6 months; window_open until expiry; grace until expiry + 6 months; lapsed afterwards. Earlier expiries follow the former regime (30 business days). null = the rule does not apply: no expiry date, neither registrada nor esperando_renovacion, or Madrid route.
milestones[] gains two derived milestones: renewal_window_open (rule: "6_months_before_expiry") and grace_end (rule: "6_months_after_expiry"), with a three-norm legalBasis: Law 19.039 art. 24 (term), Decree 82/2022 art. 27 (regulation, a new role value) and Civil Code art. 48 (computation). grace does not claim the registration is in force, and lapsed does not claim renewal was not requested. Opposition briefs do not carry renewalStatus.
What to do: nothing if unused. Read null as "unknown", not "no renewal needed", and remember the value is that of the response date.
2026-09-25 · [BEH] renewals[] publishes once a renewal act that both sources report on the same day
In the trademark record, the same renewal act reported by the Estado Diario and by the record page on the same day appeared twice in renewals[]. It is now deduplicated by (date, code) across sources, by count: if one source has 2 rows and the other 1, 2 are published. Two entries from the same source on the same day are not merged (they may be two acts), and undated entries are not deduplicated. annotations[], eventId and lastRenewalOn do not change.
What to do: nothing.
2026-09-25 · [BEH] An impossible date in trademark filters is 400, not 500
milestoneFrom/milestoneTo, filedFrom/filedTo, expiresFrom/expiresTo, registeredFrom/registeredTo and publishedFrom/publishedTo only validated the YYYY-MM-DD shape: 2026-02-30 or 2026-13-45 returned 500 internal_error. They are now validated as real calendar days (like dateFrom/dateTo of /annotations/search) and return 400 validation_error, including the year 0000. Applies to REST and MCP.
What to do: nothing; if you retried that 500, it is now a 400 that should not be retried.
2026-09-25 · [ADD] milestoneKind=opposition_close filter by opposition-window closing date
milestoneKind=opposition_close with milestoneFrom/milestoneTo (one is enough) returns trademarks whose opposition close falls in the range, in GET /trademarks, POST /trademarks/search ≡ search_trademarks (search, batch of 100, similarity, opposition window) and GET|POST /insights/trademarks ≡ get_trademark_insights. It matches exactly the opposition_close milestone published in milestones[], whether derived or estimated. No range cap. An unknown kind is still 400 validation_error.
?milestoneKind=opposition_close&milestoneFrom=2026-09-28&milestoneTo=2026-10-04
What to do: nothing if unused. For closing alerts, one call with the batch and the range is enough.
2026-09-25 · [BEH] milestoneFrom/milestoneTo without milestoneKind include opposition_close
Until now, without milestoneKind the milestone filter was equivalent to expiresFrom/expiresTo. It now matches trademarks with any filterable milestone in the range: their expiry or their opposition close. A call that already worked may return more trademarks; none starts failing. Upcoming milestones in /reports/portfolio remain expiry only.
What to do: if you call without milestoneKind expecting only expiries, add milestoneKind=expiry.
2026-09-25 · [ADD] opposition_close milestone in milestones[], derived or estimated, and legal basis in oppositionWindow
The opposition-window close becomes a milestone: {kind: "opposition_close", date, basis, anchor: {kind: "publication", date}, rule: "30_business_days_after_publication", legalBasis}. It appears in the trademark record and in every trademark of the collection, search, batch, similarity, window, list_party_trademarks and the reports. Its date is the deadline of oppositionWindow. milestones[] is ordered by date.
legalBasis cites three norms: Law 19.039 art. 5 (term), Law 19.039 art. 11 (business_days) and Law 19.880 art. 25 (computation). basis: "estimated" (new value) adds estimateReasons when the close cannot be asserted: published_on_non_business_day (published on a non-business day) or outside_holiday_calendar (period outside the declared holiday calendar; the real close falls on that day or later). No publication date, no milestone. oppositionWindow gains legalBasis and, when estimated, estimateReasons.
What to do: nothing if unused. If you show the closing date, distinguish estimated from derived.
2026-09-25 · [BREAK] A computed milestone's legalBasis changes from an object to a list of {norm, article, role}
In milestones[], the derived variant declared legalBasis: {norm, article} (an object). It is now a non-empty list of {norm, article, role} (role: term | business_days | computation), so every norm behind the term can be cited. The same applies to estimated and to oppositionWindow.legalBasis. No derived milestone had been emitted yet, so this changes the shape declared in the OpenAPI, not any response already served. expiry carries no legalBasis.
before: "legalBasis": {"norm": "Ley 19.039", "article": "5"}
after: "legalBasis": [{"norm": "Ley 19.039", "article": "5", "role": "term"}, …]
What to do: if you generated types from the OpenAPI, regenerate them and read legalBasis as a list.
2026-09-25 · [DATA] Estado Diario annotations: the text the bulletin truncated
Some bulletin entries had been stored cut short. The original bulletins were re-read and 186 Estado Diario annotations (2024-10 to 2026-08) were completed, only where the stored text was a literal prefix of the printed one. In 72 annotations observation now carries the full text (from ~2,266 to ~5,990 characters on average) and observationDerived becomes null. In one of them (id 10416963), the two versions of the event no longer match and versionsDiffer becomes true. The norm citations of those annotations are only extended.
What to do: nothing required. If you stored the observation of those annotations, re-read it: it may be longer.
2026-09-25 · [BEH] act publishes the INAPI code of the Estado Diario subsection heading
Until now act only carried the act of the A1, A3, A4, A5 and D5A headings. It now also carries the act printed as a subsection heading: under M11 "Resoluciones Varias", "Fondo-Res. Favorable Escrito" is act 664B. 84,888 annotations go from act: null to a code; 198,080 now have act in total. The same heading can be different acts depending on the section (614 in C8, 969 in C3, 966 in C2). All are INAPI codes; none are invented. actName/actLabel follow the code and are null for codes INAPI never named. As a result, 29 664C annotations enter the report's renewals[]. Applies to the collection, single annotation, cross-search, daily bulletin, trademark record and report.
What to do: if you treated act as "family A only", check the code, not the section: it can now be 664B, 428, 431… under M11, C8, M8.
2026-09-25 · [ADD] actStatus on the annotation: identified | unidentified | not_applicable
act: null could mean "the section already is the act" or "we do not know which act it is". actStatus tells them apart and is always present (never omitted or null) on every full annotation: collection, single, cross-search, daily bulletin, trademark record and report.
actStatus |
act |
meaning |
|---|---|---|
identified |
the code | act identified |
unidentified |
null |
the annotation names an act that could not be identified |
not_applicable |
null |
the section already is the act |
It does not say where the code came from. duplicates[] variants and opposition events do not carry it.
What to do: if you show act, use actStatus to decide what to show when it is null: "not applicable" or "unidentified act".
2026-09-25 · [DATA] Norms: 14 official titles, 19.039 misspellings re-pointed and anaphora fixed
Changes in /legal-norms, /legal-norms/{id} ≡ list_legal_norms/get_legal_norm and the citedNorm filter; the shape does not change.
- 14 norms get their official BCN title in
name(Law 19.039 and its Regulation, 20.254, 19.799, 19.880, among others). 19.039 changes from "Ley de Propiedad Industrial" to its official title. - The 18 "laws" that were mistyped 19.039 cease to exist and their
idreturns404(e.g.90). Their citations move to 19.039 or are dropped as duplicates. - "Del Reglamento de la citada ley" now resolves to the last law named: 239 citations move to the 19.039 regulation.
- 5 regulations with no citations and 4 citations misattributed to a "regulation 19.880" remain.
Catalogues may take up to a day to reflect this because of caching.
What to do: re-measure per-norm counts and trademarks matching citedNorm. Do not use norm ids stored before 2026-09-25.
2026-09-25 · [BEH] 27 section codes gain a name and label
27 codes get their literal INAPI name and a label in es-CL, en and pt-BR. 27,170 annotations (mostly 2012–2020) go from sectionName: null to having a name and label, and in /sections ≡ list_sections/get_section those codes gain source.label and label. The catalogue now has 673 codes, 541 named; sectionName: null still means INAPI never published the name. W177 keeps INAPI's spelling as is (Notificaci¢n Ejecutoria Rechazada). A06A does not change. Catalogues may take up to a day to reflect this because of caching.
What to do: nothing. Where you showed the raw code or "—", you will now see the name.
2026-09-25 · [BEH] carriesParties and emittedBy measured on the corpus — /sections ≡ list_sections/get_section
Both fields are filled with what the corpus supports, without overwriting existing values. emittedBy: 548 codes go from [] to a value (544 with the record page as emitter and 4 ["estado_diario"]: C9N, C12N, C14N, C16N); 99 remain []. carriesParties: C9N, C12N, C14N and C16N become true, and D5A false. null still means "unknown", never "carries no parties"; codes that come only from the record page do not get carriesParties. The shape does not change.
What to do: nothing. emittedBy: [] still means "not measured" for the remaining 99.
2026-09-24 · [BEH] The open opposition window also excludes registrada and concedida trademarks
A registered or granted trademark no longer admits opposition, even if its 30th business day after publication is still ahead. Membership of the window now also requires that its (derived) status is neither registrada nor concedida. This affects oppositionWindow: "open" in POST /trademarks/search ≡ search_trademarks, and inOppositionWindow in get_clearance_report and get_portfolio_report.
liveOnly does not change (a registered trademark is still live), nor does the trademark's oppositionWindow object (deadline/businessDaysLeft/expired): it publishes the term, not membership. So !expired no longer means "in the window"; membership is decided by oppositionWindow: "open" and inOppositionWindow.
What to do: nothing. If you showed "in window" for a registered trademark, that will no longer happen.
2026-09-24 · [ADD] Opposition briefs: GET /reports/opposition-brief ≡ get_opposition_brief and GET /reports/opposition-response-brief ≡ get_opposition_response_brief
Two new one-call reports that complete the /reports/* family.
- Filing an opposition (
applicationNumber= the target trademark, optionalownApplicationNumber,limit1–20): the trademark record withmilestones[],inOppositionWindow,similarity(only with an own trademark: the target's position among the 100 most similar, orfound: false; without an own trademark the key is omitted), the holders with their oppositions filed and received,citedNormsandcoverageFrom=2024-09-30. Non-existent own trademark ⇒404; same as the target ⇒400; outside holiday-calendar coverage ⇒422 out_of_coverage. - Answering an opposition (
applicationNumber= the opposed trademark,limit):oppositionActscollapsed by fact (up to 50 +hasMore),opponentsresolved by party identifier (never by name) andunresolvedOpponentMentions(counted, never named).
Neither returns verdicts, an opposition closing date outside milestones, or renewal status. Scopes reports:read or the brief's own (reports:opposition-brief:read, reports:opposition-response-brief:read), denied by default; one quota unit per call.
What to do: nothing if you don't use them. To use them, ask the operator for reports:read or the brief's scope.
2026-09-24 · [BEH] The open opposition window is no longer always empty
oppositionWindow: "open" in POST /trademarks/search ≡ search_trademarks always returned an empty page (estimatedTotal: 0), get_clearance_report always reported inOppositionWindow: false and get_portfolio_report always returned inOppositionWindow.trademarks: []: they required a publicada status that INAPI never emits.
Now a trademark is in the window when it has publishedOn, its 30th business day after publication is today or later (Chilean days, using the holiday calendar) and its status is not terminal. It is the same window and the same arithmetic as the trademark's oppositionWindow object, which now also counts "today" in Chilean time. If the calendar does not cover the days needed, the response is 422 out_of_coverage with details.coverage. After the change, 7,526 trademarks are in the window.
What to do: nothing in code. If you treated "empty window" as "nothing to watch", that answer was false until today.
2026-09-24 · [ADD] A party's portfolio: GET /reports/portfolio ≡ get_portfolio_report
One-call report with a party's portfolio. Input: partyId (required), relation (current by default, former, all) and horizonDays (1–366, default 90). Built from the party identifier, never from the name ("ACME" does not pull in "ACME CHILE").
Returns party, trademarkCount, byStatus[] and byClass[], upcomingMilestones (milestones between today and today + horizonDays, up to 50 + hasMore), inOppositionWindow (membership, no closing date), oppositionsReceived, representatives and relations[], with dataAsOf/coverageThrough. No renewal status and no verdicts.
Scope reports:read or reports:portfolio:read, denied by default; one quota unit per call. Non-existent party ⇒ 404 not_found.
What to do: nothing if you don't use it. To use it, ask the operator for reports:read or reports:portfolio:read.
2026-09-24 · [BEH] liveOnly excludes the seven terminal statuses (previously three)
liveOnly=true only excluded denegada, vencida and anulada. It now excludes the seven statuses that end the lifecycle: rechazada, denegada, desistida, abandonada, anulada, caducado, vencida. A trademark with no known status is still kept.
This changes everywhere liveOnly is accepted: GET /trademarks, POST /trademarks/search ≡ search_trademarks (including similarity and the opposition window), /insights/trademarks ≡ get_trademark_insights and watch: a watchlist with filters.liveOnly: true stops producing hits on trademarks in the four newly excluded statuses.
What to do: nothing if liveOnly meant "live". If you need rejected, withdrawn, abandoned or lapsed trademarks, stop sending liveOnly or filter by status.
2026-09-24 · [BEH] Similarity and opposition window: an invalid taxId or an inverted date range is 400
With similarTo/similarToTrademark or oppositionWindow: "open", a taxId not shaped like a RUT or a range with From > To (filedFrom/filedTo, expiresFrom/expiresTo, registeredFrom/registeredTo, publishedFrom/publishedTo) returned 200 with an empty list. They now return 400 with the same code as the collection, invalid_rut or invalid_range, in REST and MCP.
What to do: if you sent such values, you will now get the error instead of a misleading empty result.
2026-09-24 · [ADD] One-call reports: GET /reports/daily-bulletin ≡ get_daily_bulletin and GET /reports/clearance ≡ get_clearance_report
The /reports/* family is born, with its own scopes denied by default: reports:read grants all of them and reports:<report>:read grants one. A key with empty scopes does not get them (403 forbidden). One quota unit per call.
get_daily_bulletin(requireddate,section,limit≤ 100,cursor): what the Estado Diario published that day, row by row with parties, exacttotal,meta.bulletinsandmeta.sections(the whole day's index in three languages). A day with no bulletin is422 out_of_coveragewithcoverage: {from, through}, never an empty list.get_clearance_report(text≤ 200,niceClass1 to 10 classes): signals, never a verdict.similar[]per channel (phonetic,fuzzy,semantic) withinOppositionWindow,dominantHolders[]andcitedNorms[]. There is noregistrable/available/risk/infringesfield.
What to do: nothing if you don't use them. To use them, ask the operator for reports:read or the report's scope.
2026-09-24 · [ADD] milestones[] on the trademark and the milestoneFrom/milestoneTo/milestoneKind filters
Each trademark publishes its dated milestones in milestones[], always present ([] if there are none): in GET /trademarks/{n} ≡ get_trademark and in every trademark of GET /trademarks, POST /trademarks/search ≡ search_trademarks and GET /parties/{id}/trademarks ≡ list_party_trademarks. A milestone is source ({kind, date, basis, field}) or derived (with required anchor, rule and legalBasis). For now only expiry is published, taken from expiresOn.
New filters milestoneFrom/milestoneTo (YYYY-MM-DD, inclusive) and milestoneKind in the collection, search, batch, similarity, opposition window and /insights. milestoneFrom > milestoneTo ⇒ 400 invalid_range; milestoneKind without a range or with an unpublished kind ⇒ 400 validation_error. The watch snapshot and webhook do not change.
What to do: nothing if you don't use them (new field and parameters).
2026-09-24 · [BEH] oppositionWindow with several applicationNumber values is 400
With oppositionWindow, applicationNumber is a single portfolio trademark against which the window is ranked, not a batch. Previously, with several numbers the first was used and the rest silently ignored; the request now returns 400 validation_error, in REST and MCP.
What to do: send a single applicationNumber together with oppositionWindow.
2026-09-24 · [ADD] A party's relations: GET /parties/{id}/relations ≡ list_party_relations
New route and tool, with the common collection envelope (no total) and scope brands:read. Input: type (list; all seven by default), relation (current, former, all), limit ≤ 100 and cursor. Non-existent party ⇒ 404.
Seven edge types, built from party identifiers: co_holder, represents/represented_by, assigned_to/assigned_from (A07T transfers, since 2011-08-01) and opposed/opposed_by (since 2024-09-30). Each edge carries label, party, trademarkCount, firstOn/lastOn, up to 5 numbers in evidence, evidenceTruncated and coverageFrom. Opponents that cannot be identified produce no edge: they are counted in meta.unresolvedOpponentMentions. There are no "same group" or name-similarity edges: these are signals, not verdicts.
What to do: nothing if you don't use it.
2026-09-24 · [BEH] /parties ≡ list_parties: meta.coverageThrough carries a value and hasMore is correct with limit=100
In GET /parties ≡ list_parties:
meta.coverageThroughcarries its date (previously alwaysnull).- With
limit=100and more than 100 parties, the response saidhasMore: falseandnext: null; it now sayshasMore: trueand returns the cursor. sort=namereturns the same results, but much faster with any filter.
What to do: nothing. If you treated /parties meta.coverageThrough as always null, you will now get a date. To walk the full register, use sort=name.
2026-09-24 · [ADD] /insights: groupBy=expiresMonth (with meta.undatedCount) and groupBy=transferYear
Two new groupBy values in GET /insights/trademarks, POST /insights/trademarks/search and get_trademark_insights, with the same { key, label, count } shape.
expiresMonth: trademarks by expiry month;key= first day of the month (2027-03-01),labelin three languages. Trademarks without an expiry date fall in no month and are declared inmeta.undatedCount(present only with thisgroupBy, also when0): Σcount+undatedCount= filtered trademarks.transferYear: by year of the act, trademarks with a transfer (A07T) that removed a holder. A trademark counts in every year it had one. Coverage since 2011-08-01.
What to do: nothing if you don't use them. An exhaustive switch over groupBy or over meta keys receives new values.
2026-09-24 · [BEH] /insights?groupBy=niceClass: the class label comes in all three languages
In GET /insights/trademarks, POST /insights/trademarks/search and get_trademark_insights, each class label changes from {"es-CL":"Clase 35","en":null,"pt-BR":null} to {"es-CL":"Clase 35","en":"Class 35","pt-BR":"Classe 35"}. It is the labelled number, not the class heading (which lives in /nice-classes). Same shape; a null becomes a string.
What to do: nothing.
2026-09-24 · [BREAK] MCP: the twin tools list_trademarks and search_trademark_insights are withdrawn
Only MCP changes; REST does not. The server goes from 33 to 31 tools. The two withdrawn tools were exact copies of another tool (same input schema and same response):
| withdrawn | use instead |
|---|---|
list_trademarks |
search_trademarks (including the applicationNumber batch, with meta.notFound) |
search_trademark_insights |
get_trademark_insights |
Calling a withdrawn name returns an unknown-tool error. The REST routes GET /trademarks and POST /insights/trademarks/search remain. The availability_check prompt now uses search_trademarks. The published MCP manifest announces the prompts the server actually serves (availability_check, case_history, weekly_watch) and its four resources.
What to do: in MCP, replace list_trademarks with search_trademarks and search_trademark_insights with get_trademark_insights, with the same arguments. In REST, nothing.
2026-09-24 · [BEH] The contract date moves to 2026-09-24
The version declared by the OpenAPI document (info.version), the MCP server (serverInfo.version) and the manifest moves from 2026-09-23 to 2026-09-24, the effective date of this day's changes. Same shape, different value.
What to do: nothing, unless you compare the version against a fixed value.
2026-09-24 · [BEH] Similarity and the opposition window apply filters they used to ignore; sort and unsupported combinations are 400
With similarTo/similarToTrademark in POST /trademarks/search ≡ search_trademarks, the filters holder, representative, taxId (with relation), the four date pairs, registrationNumber and liveOnly are now applied; before, they were accepted and ignored. niceClass with several classes returns the union (previously an empty list). oppositionWindow: "open" applies the same filters and accepts several classes.
These are now 400 validation_error: sort other than relevance with similarity or opposition; denomination and facets with similarity or opposition; allClasses with similarTo or oppositionWindow. Watch does not change.
What to do: check whether you sent those filters (they now really filter) or those combinations (they are now errors).
2026-09-24 · [BEH] GET /trademarks publishes an exact total when possible; otherwise the key is omitted (previously null)
Trademark collections sent "total": null when the total was not exact. Now, like every other collection:
- Without filters,
totalis the exact number of trademarks in the corpus, while it remains exact; otherwise it is omitted. - With filters and up to 10,000 results, the exact count is published.
- With more than 10,000, with
q, with similarity or with the opposition window, the key is omitted.
meta.estimatedTotal does not change. The applicationNumber batch already published its exact total. REST ≡ MCP.
What to do: treat total as optional; do not expect null.
2026-09-24 · [DATA] /nice-classes moves to the WIPO 13-2026 headings: title.en and title.pt-BR are no longer null
GET /nice-classes, list_nice_classes and the iris://catalogs/nice-classes resource serve the headings of the Nice Classification, edition 13-2026 (applied in Chile since 2026-01-01). The shape does not change; the title values do:
title.es-CL: WIPO's official Spanish translation (previously INAPI's online classifier text, from earlier editions). Classes 1, 10, 35, 36, 40, 42 and 45 change.title.en: WIPO's authentic text (previouslynull).title.pt-BR: Tarno's own translation, not official (previouslynull).
What to do: nothing in code. If you stored the previous title.es-CL, you will see different text for those seven classes. If you show pt-BR to a lawyer, say it is a translation.
2026-09-23 · [BREAK] Single envelope for trademark collections, unknown parameters → 400, dedupe by default and the goodsAndServices filter
The change that closes the contract cut, in REST and MCP at once:
GET /trademarks,POST /trademarks/search, the batch andGET /parties/{id}/trademarksreturn the common envelope{ results, limit, next, hasMore, meta }, and no longer sendETagorIris-Data-As-Of(freshness is inmeta.dataAsOf/meta.coverageThrough).- Facets,
matchedFields[].field,highlight.fieldandkindSourceuse English names. - Any undeclared query parameter, body key or MCP argument is
400 validation_error(code: "unrecognized_keys"). GET /trademarks/{n}collapses by event by default (dedupe=true);GET /trademarks/{n}/annotationsacceptsdedupe.- The batch applies the other filters;
meta.notFoundlists only what does not exist. /insightsfillsmeta.dataAsOfandmeta.coverageThrough.- Catalogs send catalog
Cache-Controlon 2xx responses only. - New filter
goodsAndServices(goods-and-services text, 2–200 characters). sort=statussorts by the publishedstatus.
The full before→after map is in the migration guide.
What to do: follow the migration guide: read trademark collections from results/meta and remove undeclared parameters.
2026-09-23 · [DATA] Merge of duplicate parties: 20,731 groups
On 2026-09-09 duplicate parties in the corpus were merged: 20,731 groups, with 25,152 party records absorbed into the party that is kept. The earlier record only covered a previous pass of 179 groups.
What to do: if you stored party identifiers or party counts obtained before 2026-09-09, query them again.
2026-09-23 · [BEH] GET /trademarks and POST /trademarks/search reject citedNorm/article/letter instead of ignoring them
GET /trademarks?citedNorm=19039 (or search_trademarks with that argument) returned 200 with the first page unfiltered. It now returns 400 validation_error naming the field, in REST and MCP: citedNorm, article and letter filter annotations by the norm they cite, not trademarks.
What to do: to filter by cited norm, use POST /annotations/search ≡ search_annotations.
2026-09-23 · [BREAK] status carries the derived status; INAPI's literal travels in statusInapi; derivedStatus is withdrawn
INAPI writes the registration number and publishes the grant before updating its status. status now publishes the derived status: a trademark INAPI still calls en_tramite (or another pending status) becomes registrada if it has a registration number or date, and concedida if it has a grant annotation or a granted class. Otherwise it is INAPI's literal.
statusInapi: INAPI's literal only when it differs;nullwhen it matches. Always present.statusStale: in every trademark view;trueif INAPI has not updated it yet.derivedStatus: withdrawn.- The
statusfilter,groupBy=statusand thestatusfacet operate on the derived value.
Applies to every trademark view in REST and MCP. Earlier watch snapshots keep their status with statusInapi and statusStale set to null. About 660 trademarks (0.08%) are affected.
What to do: read the status from status (not derivedStatus), use statusStale and statusInapi to show that INAPI has not updated it, and review any fixed per-status counts. Details in the migration guide.
2026-09-23 · [BEH] /insights applies the full trademark filter or returns 400
GET /insights/trademarks, POST /insights/trademarks/search and get_trademark_insights accepted the full trademark filter but applied only nine fields and, for niceClass/status/signType, only the first value: filtered figures published until now were wrong.
Now q (exact lexical match, no cap), liveOnly, relation, registrationNumber, applicationNumber (batch, max 100), the date ranges and all values of niceClass, status and signType are applied. These return 400 validation_error: similarTo, similarToTrademark, allClasses, oppositionWindow, citedNorm, article, letter, facets and minScore.
What to do: request again any insights figure obtained with filters.
2026-09-23 · [BEH] POST /annotations/search no longer times out, versionsOfFact counts the whole fact and four filters are applied
In POST /annotations/search ≡ search_annotations:
- Searching without filters (or with
citedNorm) no longer returns504. Cursors issued before remain valid. versionsOfFactcounts every version of the fact, not only those passing the filter.dateFrom/dateTo(YYYY-MM-DD, inclusive),niceClass(1–45) anddedupeare applied; previously they were dropped without error.
dedupe defaults to true: each fact appears once, with its other versions in duplicates[]. Page lengths change.
What to do: if you need each version as its own row, send dedupe: false. Date or class filters you were already sending start filtering.
2026-09-23 · [BREAK] The seven labels still in Spanish move to English
Rename only, in REST and MCP: same data and types. eventoId → eventId (annotations); categoria/categoriaEtiqueta → category/categoryLabel and ambito/ambitoEtiqueta → scope/scopeLabel (GET /trademarks/{n} ≡ get_trademark); titular → holders in each trademark of the collections and in the snapshot of watch hits; observacion → observation in the embedded party; filters.clase → filters.niceClass in watchlists. The embedded party's kind becomes natural/legal/unknown, as in /parties.
Spanish values do not change. The embedded party now includes observation, which previously never arrived. Existing watchlists do not need to be recreated.
What to do: rename these keys when reading and write filters.niceClass in watchlists (filters.clase is now an error). Full map in the migration guide.
2026-09-23 · [BREAK] /holidays and /system/coverage filter by from/to; the Spanish names are 400
GET /holidays ≡ list_holidays and GET /system/coverage ≡ get_system_coverage only understood desde/hasta (and anio/tipo) and ignored from/to, returning the whole calendar or the whole coverage.
before: ?desde=…&hasta=…&anio=…&tipo=… meta.coverage { desde, hasta }
now: ?from=…&to=…&year=…&type=… meta.coverage { from, to }
The old names and any other unknown parameter return 400 validation_error; the detail gives the new name (e.g. `desde` was renamed: use `from`).
What to do: rename the four parameters and read meta.coverage.to instead of meta.coverage.hasta.
2026-09-23 · [BEH] GET /parties/{id}/trademarks no longer always returns 400
The route always returned 400 validation_error (id: expected string) unless the identifier was repeated as ?id=. Now id travels only in the path and the query string accepts role, relation, limit and cursor. The list_party_trademarks tool already worked and does not change.
What to do: nothing; if you added ?id=, remove it.
2026-09-23 · [BEH] Annotations with parties no longer return 500
Every response that included an annotation with parties failed with 500: GET /annotations/{id}, GET /trademarks/{n}/annotations, GET /trademarks/{n}?expand=annotations and their twins get_annotation and list_trademark_annotations. It affected 462,439 annotations on 187,886 trademarks. They are now delivered in the shape the contract publishes: parties[] with role, name, method, candidates and partyId.
What to do: nothing if you already read those fields as the contract says.
2026-09-23 · [BREAK] The published OpenAPI no longer declares /watchlists/ with a trailing slash
The generated OpenAPI document declared the watch collection as /watchlists/; it now says /watchlists, like every other route. The server always answered at /watchlists, so no request changes; only the path a client generator reads changes.
What to do: if you generated a client from the OpenAPI document, regenerate it.
2026-09-23 · [BREAK] The cut: /v1 stops responding
The /v1 prefix stops responding: no redirect, no grace period and no version header. An old URL returns 404. Routes, field names, collection envelope, pagination, errors and MCP tools change at once; there is no period in which both shapes coexist.
Watch out for RateLimit-Reset, which replaces X-Quota-Reset: it is a number of seconds remaining, not an absolute instant. Renaming the header without changing the calculation produces absurdly long waits.
Code values (still in Spanish: status: "en_tramite"), natural identifiers and corpus data do not change.
What to do: migrate in one go following the migration guide, which has the full before→after map of routes, fields and tools.
2026-09-23 · [BREAK] The contract date becomes the effective date: 2026-09-23
The version declared by the OpenAPI document, the MCP server and the manifest moves from 2026-09-20 to 2026-09-23. From now on the contract date is the effective date: it identifies the contract being served, not the day it was designed, and it moves whenever the surface changes.
What to do: if you compare the version against a fixed value, update it to 2026-09-23.
2026-09-22 · [BEH] liveOnly=true excludes vencida trademarks again
When field names were renamed to English, a code value was changed by mistake: liveOnly=true excluded an expired status that does not exist, so trademarks in the terminal status vencida came back as "live". vencida is restored. Code values stay in Spanish; only field names changed.
The response shape is unchanged. With liveOnly=true, expired trademarks are actually excluded again.
What to do: nothing. If you added your own filter for expired trademarks as a workaround, you can remove it.
2026-09-22 · [BREAK] sort tokens are published in English only
The sort tokens still accepted Spanish aliases next to the English ones. The contract now uses one language. relevancia (→ relevance), recientes (→ filedOn:desc) and antiguas (→ filedOn:asc) are removed. estado, estado:asc and estado:desc become status, status:asc and status:desc. The published OpenAPI document shows the new list.
What to do: replace the Spanish tokens with their English equivalents. The full table is in the migration guide.
2026-09-22 · [BEH] Encumbrances return data again, and territory.regionCode is no longer empty
Because of a bug in the field rename, encumbrances[] always came back as [], even when the trademark had registered encumbrances. It now returns the real ones. The same bug left territory.regionCode without a value in the parties resource. That is fixed too.
What to do: nothing. If you cached responses with encumbrances: [] or an empty regionCode, fetch them again.
2026-09-22 · [BREAK] Sorting by fechaActualizacion is removed
fechaActualizacion:desc and fechaActualizacion:asc are gone. They sorted by the date our ingestion last touched the record, not by an INAPI fact, and the contract does not publish that field. The other ten sorts stay: filedOn, denomination, status, expiresOn, registeredOn and publishedOn, each with asc|desc. Each one has guaranteed performance.
What to do: if you used fechaActualizacion, switch to another published sort. To track recent changes, use watchlists. Details are in the migration guide.
2026-09-22 · [BREAK] The OpenAPI document is generated from the schemas, versioned by date and declares the domain
The published OpenAPI document is no longer written by hand. It is generated from the same schemas the server uses. You will see three changes:
info.versionmoves from1.0.0to the contract date (2026-09-20), the same version the MCP server declares. The contract is versioned by date, and its history is in this changelog.servers[0].urlishttps://api.tarno.cl.- No SDK is published. The document is the contract, and you can generate your own client from it.
The document no longer lists GET /trademarks/{applicationNumber}/encumbrances (request encumbrances with expand=encumbrances) and adds GET /system/vigilance-horizon.
What to do: if you compared info.version against a semantic version, compare the date instead. Point your clients at api.tarno.cl.
2026-09-22 · [ADD] GET /system/me, ETag caching and the public trial key
These changes are additive:
GET /system/me≡get_system_mereturns your identity (organization), your scopes, the quota you have used, your limit and when the window resets. It only shows your own data, and the response is sent withno-store.- Single objects carry an
ETag. If you repeat a read withIf-None-Matchand nothing changed, you get304with no body. Catalogs are served with long cache lifetimes. - Public trial key: 1,000 requests per month and no watchlist scopes (neither
watch:readnorwatch:write). Moving to a paid key is handled manually with Tarno.
What to do: nothing. Optionally, send If-None-Match so you do not download data that has not changed.
2026-09-22 · [BREAK] MCP tools are renamed after their routes, and resources, prompts and a version are added
MCP tool names now match the new contract. There are no aliases and no transition period. Some examples: search_brands → list_trademarks, find_similar_brands → search_trademarks with similarTo, get_brand_detail/get_brand_report → get_trademark (with expand=), get_brand_opposition → list_trademark_annotations with family=opposition, watch_* → create_watchlist, list_watchlists, get_watchlist, update_watchlist, delete_watchlist…, and list_system_coverage → get_system_coverage.
New: the catalogs as resources (iris://catalogs/states, iris://catalogs/sections, iris://catalogs/nice-classes, iris://catalogs/holidays/{year}), three prompts (availability_check, case_history, weekly_watch), the explain_section_code tool (it says "unknown" when a code has no legend), and a server version equal to the contract date (2026-09-20), with the four usage rules in its instructions.
What to do: update the tool names using the MCP map in the migration guide.
2026-09-22 · [BREAK] Watchlists move to /watchlists, with the common envelope and a fixed webhook body
- The
/v1/watch/*routes are replaced by/watchlists,/watchlists/{id},/watchlists/{id}/items,/watchlists/{id}/hitsand/watchlists/{id}/deliveries. - Watchlist collections no longer return bare arrays or
{items,nextCursor}. They use the common envelope (results[],limit,next,hasMore,total,meta). - The
sinceparameter is replaced by the common cursor. - The webhook body is now fixed:
{ watchlistId, sweepAt, newHits, fetchUrl }, signed withX-Tarno-Signature(HMAC-SHA256). Any future change to it will be announced as a[BREAK]. - An ID that belongs to another tenant still gets the same response as an ID that does not exist:
404.
What to do: migrate your routes, envelope parsing and pagination using the migration guide, and verify the webhook signature.
2026-09-22 · [ADD] Day-by-day coverage, the watchlist horizon and an ID for each encumbrance
GET /system/coverage≡get_system_coverageshows, day by day, which bulletins were processed and how many acts each one brought, with holidays marked. You can tell a day with no bulletin because of a holiday from a real gap, and an undeclared year from a year without problems.GET /system/vigilance-horizon≡get_vigilance_horizontells you how far back you can ask about changes (2 months today).- Each encumbrance has a stable ID, derived from the annotation that records it. You can now tell apart two encumbrances of the same type on the same trademark.
What to do: nothing. These are new features.
2026-09-22 · [ADD] Statute citations and the section catalog are published
- Each annotation now lists the statutes it cites, with article, letter and paragraph. The
GET /legal-normsandGET /legal-norms/{id}catalog (≡list_legal_norms/get_legal_norm) returns each statute with its articles inside the object. Annotations and trademark search accept thecitedNorm,articleandletterfilters. GET /sectionsandGET /sections/{code}(≡list_sections/get_section) list the section codes with their family, sources, whether they carry parties, and their confirmed equivalences. A code with no legend publishes its label asnull, never an invented name.
What to do: nothing. These are new features.
2026-09-22 · [BEH] Annotations are grouped by fact by default, and a flag shows when versions differ
Before, a fact published by two sources came back once per copy unless you sent dedupe=true. Now each fact comes back once by default. To see the copies, send dedupe=false. The copies also stay in duplicates[], so grouping never deletes anything.
Every annotation now always has two new fields: versionsOfFact (how many rows describe the same fact) and versionsDiffer (true when another version is neither identical to this one nor contained in it). There are 593 pairs that say different things, out of 580,218 facts. The flag does not say which source published which version. On former parties, retiredBy becomes a neutral code (authority_withdrew) and keeps INAPI's section codes. renewals and lastRenewalOn give the same result with either value of dedupe.
What to do: if you need every copy, send dedupe=false. Check versionsDiffer.
2026-09-22 · [BREAK] Field names move to English, and labels come in three languages
Labels (seccionEtiqueta, seccionActoEtiqueta, categoriaEtiqueta, ambitoEtiqueta and the renewal label) are no longer chosen with Accept-Language. Each one is an object with all three languages at once (es-CL, en, pt-BR), with null where there is no translation. The response is the same for any Accept-Language.
About 180 field names move to English. For example:
nroSolicitud → applicationNumber denominacion → denomination
estado → status fechaPresentacion → filedOn personas → parties
anotaciones → annotations variantes → duplicates soloVigentes → liveOnly
Some things do not change: code values (status is still "en_tramite"), INAPI literals (section codes, annotation text, RUT), and categoria/ambito.
What to do: rename the fields using the full map in the migration guide, and read labels as three-language objects.
2026-09-21 · [BREAK] Persons become parties, a single entity with a role
GET /v1/personas and search_personas are replaced by GET /parties, GET /parties/{id} and GET /parties/{id}/trademarks (≡ list_parties, get_party, list_party_trademarks). Holder, applicant, representative and opponent are now one entity with a role, so the same party no longer appears several times.
Shape changes: tipo/tipoOrigen → kind/kindSource (the second can now be null), revision → review, with its reason as a code, region/comuna → territory (regionCode, regionName, communeCode, communeName, unknownCommune; null for foreign parties), and marcasTotal → trademarkCount. The 25,645 parties with no trademarks still appear, with a count of zero.
What to do: migrate to /parties using the map in the migration guide.
2026-09-21 · [ADD] GET /nice-classes: the Nice classification catalog
A new catalog with the 45 Nice classes and their headings, as published in INAPI's classifier. GET /nice-classes ≡ list_nice_classes. You can cache it for one year.
What to do: nothing. This is a new feature.
2026-09-21 · [BREAK] The holiday calendar moves to a single route, /holidays
/v1/feriados and /v1/feriados/cobertura merge into GET /holidays, and get_feriados + get_feriados_cobertura merge into list_holidays. The declared coverage always comes in meta.coverage. Asking for a date outside the coverage still returns 422 out_of_coverage.
What to do: use /holidays and read the coverage from meta.coverage. Details are in the migration guide.
2026-09-21 · [BREAK] Freshness and sync runs move to /system
/v1/freshness and /v1/sync-runs move to GET /system/freshness and GET /system/sync-runs. The get_freshness and get_sync_runs tools are renamed get_system_freshness and list_sync_runs. This keeps operational data separate from the business catalog.
What to do: update routes and tool names using the migration guide.
2026-09-21 · [BREAK] Statistics go from five routes to one with groupBy
The five /v1/insights/* routes and their five tools are replaced by GET /insights/trademarks?groupBy=… and POST /insights/trademarks/search (≡ get_trademark_insights / search_trademark_insights). The groupBy values: filings-over-time → filedMonth, by-estado → status, by-class → niceClass, by-holder → holder, by-representante → representative.
All of them return { key, label, count } items inside results[]. label includes all three languages (es-CL, en, pt-BR).
What to do: replace each route with its groupBy value using the migration guide.
2026-09-21 · [BREAK] Annotations reach MCP, and get_brand_opposition is removed
MCP now has the same three ways to request annotations as REST: get_annotation (≡ GET /annotations/{id}), search_annotations (≡ POST /annotations/search) and list_trademark_annotations (≡ GET /trademarks/{n}/annotations). get_brand_opposition is removed. Use list_trademark_annotations with family: "opposition" instead. The filters are the same in both transports.
Fix: all three forms ignored offset and returned the first page with 200. An offset now returns 400 validation_error in REST and MCP.
What to do: replace get_brand_opposition and paginate with the cursor. See the migration guide.
2026-09-21 · [BREAK] The six MCP trademark tools become three
search_brands and get_brands_batch → list_trademarks. Request a batch with applicationNumber: [a, b, c] (100 at most). It returns { items, notFound }, and an empty batch returns an empty batch. find_similar_brands and get_open_oppositions → search_trademarks with similarTo / similarToTrademark or oppositionWindow: "open". You cannot combine them (400 validation_error). get_brand_detail and get_brand_report → get_trademark with expand (classes, parties, annotations, encumbrances, norms). null means "not requested" and [] means "there are none".
What to do: update your calls using the MCP map in the migration guide.
2026-09-21 · [BREAK] The opposition window and the freshness note move
The data that came with the opposition timeline moves as follows. ventanaOposicion moves to the trademark object (GET /trademarks/{applicationNumber}) and is always present. It is null when the trademark has no publication date, which does not mean "no opposition". freshnessNote moves to meta.freshnessNote, only when you request annotations of the opposition family. It warns that a recent opposition may not appear yet.
Also, meta.coverageThrough on the annotations collection, which always came back null, now has the real date.
What to do: read the window from the trademark object and the note from meta. See the migration guide.
2026-09-21 · [BREAK] /v1/brands and /v1/oppositions no longer exist
All of their routes are removed. Points to check when you migrate:
- Similarity is two filters:
similarTo(free text) andsimilarToTrademark(excludes the source trademark and stays within its classes;allClasseswidens the search). relation(current/former/all, the party link) is not the same asliveOnly(excludes trademarks in a terminal status).holder,representativeandtaxIdaccept several values, combined with OR.- An empty batch returns an empty batch. If you go over the cap, the response is still
422 batch_too_large. - The OpenAPI document now lists the error responses for every route.
What to do: migrate using the route map in the migration guide.
2026-09-21 · [BREAK] Annotations become a resource with an ID and no source
There are three new ways to request annotations: GET /annotations/{id}, GET /trademarks/{n}/annotations?family=opposition (replaces GET /v1/brands/{n}/opposition) and POST /annotations/search. They are sorted by date (oldest first, undated last), and the cursor is composite.
Each opposition act now has its ID, so you can request it, cite it or match it with its statute citations. The fuente field leaves the contract: it told you how we ingested the record, which is not an INAPI fact.
What to do: stop reading fuente and use the new routes from the migration guide.
2026-09-21 · [BREAK] Trademarks go from six routes to four forms
GET /v1/brands → GET /trademarks. The batch → GET /trademarks?applicationNumber=a,b,c (100 at most). Similarity and open oppositions → POST /trademarks/search with similarTo or oppositionWindow: "open". Detail and report → GET /trademarks/{applicationNumber}?expand=parties,annotations,encumbrances,classes,norms. In expand, [] means "there are none" and null means "not requested".
There is one set of English filters for /trademarks, /trademarks/search and /insights. relation=current|former|all replaces soloVigentes. minScore, denomination, registrationNumber, expiresFrom/To, registeredFrom/To, publishedFrom/To and facets are kept. oppositionWindow: "closed" and "any" return 422 opposition_window_unsupported.
What to do: migrate using the route map in the migration guide.
2026-09-21 · [BREAK] Trademarks use one rule for "no value": null
The detail view published null for fields with no value, while the report left them out. Now both views always include eleven fields, set to null when there is no value: nroRegistro, estado, tipoSigno, fechaPresentacion, fechaVencimiento, fechaPublicacion, fechaRegistro, traduccion, descripcionEtiqueta, protectionDescription and imagenUrl. The exception: categoria, ambito and their labels are still left out when they are unknown. INAPI's '0' placeholder for "no number" is still not published: it now comes out as null.
What to do: if you checked whether the key existed, check for null instead. Details are in the migration guide.
2026-09-21 · [BREAK] Errors switch to RFC 9457, and three codes are renamed
All errors, in REST and MCP, use application/problem+json with type, title, status, detail, code, requestId and errors[]. The old body was { error: { code, message, requestId } }. A 400 now names the failing parameter in errors[]. Status codes are fixed: 400 = validation, 404 = does not exist or is not yours, 422 = business rule. A 5xx never includes internal details.
Renamed codes: fuera_de_cobertura → out_of_coverage, cobertura_no_declarada → coverage_not_declared, invalid_input → validation_error (MCP), and invalid_input → batch_too_large (batch cap, now 422 in MCP too).
What to do: update your error handling for the new body and codes. See the migration guide.
2026-09-21 · [BREAK] Headers: correlation, standard quota names, ETag and version
X-Request-Id → X-Correlation-ID. You can send your own, and it comes back in every response, MCP included. X-Quota-Limit/-Remaining/-Reset → RateLimit-Limit/-Remaining/-Reset. Note that RateLimit-Reset changes from an absolute time to the number of seconds left. The x-ratelimit-* headers are gone, and Retry-After is unchanged.
Single objects carry ETag (If-None-Match → 304), plus Iris-Data-As-Of and Iris-Coverage-Through. Catalogs declare Cache-Control. Watchlists publish a version. If you send it with an edit and it is no longer current, you get 409 concurrent_modification with currentVersion. If you leave it out, edits work as before.
What to do: rename the headers and recompute your wait from RateLimit-Reset. See the migration guide.
2026-09-21 · [BREAK] One envelope for every collection, and one way to paginate
Every collection returns the same shape:
{ "results": [ … ], "limit": 20, "next": "…", "hasMore": true, "total": 14,
"meta": { "dataAsOf": "…", "coverageThrough": "…" } }
It replaces data, items, nextCursor and bare arrays. total is only included when it is exact. The search estimate goes in meta.estimatedTotal. ?envelope= is removed: no parameter changes the type of a response.
Pagination is limit + cursor, and offset leaves the contract. The cursor is opaque, so do not build it yourself. Deep sorts have a cap. A cursor from a different sort falls back to page one.
What to do: read results[] and paginate with next. See the migration guide.
2026-09-18 · [BREAK] The report no longer publishes the renewal chain (renovadaDe, renovadaPor)
renovadaDe and renovadaPor are removed from the registrability report (REST and MCP). They said which application renewed which, but INAPI renews on the same application, and no current source carries that chain. It only existed for 606 trademarks (0.07%), carried over from a historical load.
before: { "renovadaDe": "1700100", "renovadaPor": "1700200", "renovaciones": [ … ] }
after: { "renovaciones": [ … ], "ultimaRenovacion": "2026-09-17" }
What to do: use renovaciones[] and ultimaRenovacion (now renewals[] and lastRenewalOn). See the migration guide.
2026-09-18 · [BREAK] The report replaces free text with code and label, and publishes renewal acts
In the registrability report (REST and MCP), tipoNombre and subtipoNombre, which mixed three dimensions, are removed. In their place come categoria (7 codes, such as marca or frase_propaganda) and ambito (5 codes, such as productos_servicios), each with its label (categoriaEtiqueta, ambitoEtiqueta). They are left out when unknown. An invented value returns 400.
The report also adds renovaciones[] (always present) and ultimaRenovacion. Each act has anotacionId, fecha, codigo, nombre and etiqueta. There are 17 codes, covering 693,357 annotations on 211,921 trademarks (25.1% of the corpus). Codes that deny a renewal are excluded.
What to do: read the codes instead of the free text, and use renovaciones[]. The current names are in the migration guide.
2026-09-17 · [BREAK] apellido is removed from personas[] and from the holder and representative statistics
The apellido field is removed from parties in the trademark detail and report, and from the insights/by-holder and insights/by-representante rows, in REST and MCP. It was always null (0 values across 397,878 persons). INAPI never publishes names split into given name and surname, and the full name is in nombre.
before: { "id": "42", "nombre": "ESTUDIO CAREY LIMITADA", "apellido": null, "pais": "CL" }
after: { "id": "42", "nombre": "ESTUDIO CAREY LIMITADA", "pais": "CL" }
What to do: nothing, unless your types declare apellido as required. In that case, remove it.