Skip to content

Analysis and calculation

Neutral tools that make a calculation or gather signals over the corpus and return it with its source: which data they used, under which rule and as of which date. They never give a verdict — neither "this mark can be registered" nor "this term ends on that day" without the article behind it. The legal decision belongs to the reader.

The complete, always up-to-date per-field schema lives in the interactive OpenAPI reference at /docs (Scalar). This guide explains the contract and links to that reference — it does not restate the schema. When a field is not documented here, /docs is authoritative.

Throughout the guide the base URL is written as $TARNO_BASE_URL and the key as $TARNO_API_KEY, and every call carries the X-API-Key header. Each REST operation has a twin MCP tool that returns the same JSON, byte for byte.

Business-day calculator

GET /business-days · scope brands:read · MCP equivalent: compute_business_days

Counts N business days strictly after a date and says everything it used to count them. It is the same count the corpus uses for a trademark's opposition window: Monday to Friday, without the holidays of the calendar, starting the day after the anchor.

curl -sS "$TARNO_BASE_URL/business-days?from=2026-09-14&businessDays=5" -H "X-API-Key: $TARNO_API_KEY"
{
  "anchor": { "date": "2026-09-14", "kind": null, "onBusinessDay": true },
  "businessDays": 5,
  "date": "2026-09-22",
  "term": { "source": "caller", "preset": null, "legalBasis": null },
  "legalBasis": [
    { "norm": "Ley 19.039", "article": "11", "role": "business_days" },
    { "norm": "Ley 19.880", "article": "25", "role": "computation" }
  ],
  "calendar": { "table": "feriado", "coverage": { "from": "2013-01-01", "through": "2030-12-31" } },
  "skipped": {
    "weekendDays": 2,
    "holidays": [{ "date": "2026-09-18", "name": "Independencia Nacional", "type": "regular" }]
  },
  "basis": "derived",
  "todayChile": "2026-09-28",
  "elapsed": true,
  "businessDaysLeft": 0
}
Parameter What it does
from The anchor date (YYYY-MM-DD, a real date: 2026-02-31 is a 400). Counting starts the next day.
businessDays How many business days to count (1…250). You set the number.
preset A term with its article, by name. Today only opposition.
anchorKind A free label for the anchor (≤ 60 characters, e.g. notification); echoed in anchor.kind, it does not change the computation.

Send exactly one of businessDays or preset: both, or neither, is a 400 validation_error that says so, the same over REST and MCP.

A number is not a term — term

Counting 10 business days is arithmetic; saying that a legal term ends that day requires the norm that sets the term. The answer keeps the two apart:

  • legalBasis ALWAYS cites the norms of the counting method: Ley 19.039 art. 11 (the terms of the law are in business days) and Ley 19.880 art. 25 (which days are non-business days and from which day the count starts).
  • term says where N comes from. With businessDays, term.source is caller and term.legalBasis is null: you chose the 10, and the system does not claim it is a legal term.

The only preset: opposition

preset=opposition counts 30 business days from the publication of the extract (Ley 19.039 art. 5). With it, term.source is preset, term.legalBasis cites art. 5 and legalBasis carries the three articles — the same legal basis and the same date as the opposition close of a trademark's record (opposition_close in milestones).

curl -sS "$TARNO_BASE_URL/business-days?from=2026-09-14&preset=opposition" -H "X-API-Key: $TARNO_API_KEY"
{
  "anchor": { "date": "2026-09-14", "kind": "publication", "onBusinessDay": true },
  "businessDays": 30,
  "date": "2026-10-28",
  "term": {
    "source": "preset",
    "preset": "opposition",
    "legalBasis": [{ "norm": "Ley 19.039", "article": "5", "role": "term" }]
  },
  "legalBasis": [
    { "norm": "Ley 19.039", "article": "5", "role": "term" },
    { "norm": "Ley 19.039", "article": "11", "role": "business_days" },
    { "norm": "Ley 19.880", "article": "25", "role": "computation" }
  ],
  "skipped": { "weekendDays": 12, "holidays": [ … "2026-09-18" …, … "2026-10-12" … ] },
  "basis": "derived",
  …
}

Why are there no other presets? Because a preset is a promise that the term has an article, and today only the opposition term is verified. The 30-, 10- or 60-day terms that accompany other INAPI decisions (substantive objections, for instance) have no verified norm that sets them, so they are not offered by name: if you need them, count them with businessDays and the answer will say you chose N. No extension of terms by exempt resolution is applied either.

When the date is an estimate — basis

basis: "derived" means the date can be stated. basis: "estimated" comes with estimateReasons, which says why not:

reason what happened what it means
anchor_on_non_business_day the anchor falls on a Saturday, Sunday or holiday the law does not say from which day to count then; the count starts the next day anyway
outside_holiday_calendar the span [anchor, date] leaves the calendar's declared coverage (calendar.coverage) weekends and only the holidays the calendar knows were skipped: the real date is the same day or later, never earlier

An estimate is presented as an estimate. Never as a derived date.

The rest of the answer

  • skipped — weekendDays are the Saturdays and Sundays skipped; holidays the holidays that fell on a weekday, with their name and type. A holiday on a Saturday or Sunday counts as weekend. It always holds: calendar days between the anchor and the date = N + weekendDays + skipped holidays.
  • calendar — the table used (feriado) and the coverage the calendar claims (null if it declares none). See coverage is finite.
  • todayChile, elapsed and businessDaysLeft — evaluated today, Chile time: elapsed is todayChile after date; businessDaysLeft counts the business days of the count that have not passed yet (0 once elapsed). That is why this answer is not cached.

Over MCP

{ "name": "compute_business_days", "arguments": { "from": "2026-09-14", "businessDays": 5 } }

Returns the same JSON as GET /business-days. An agent should state the date with its article when it uses the preset, and say it is a count when N was set by whoever asked.

Permission

brands:read, the same as the trademark reads: a key with no scope restriction and the trial key have it. A key without it gets 403 forbidden naming brands:read, before any quota is consumed — over REST and MCP (see Authentication & scopes).

Deadline to appeal a rejection or a partial acceptance

GET /trademarks/{applicationNumber}/deadlines · scope brands:read · MCP twin: get_deadlines

The last day to appeal a REJECTION (024) or a PARTIAL ACCEPTANCE (444) before the Tribunal de Propiedad Industrial: the 15th business day STRICTLY after that resolution was published in the Estado Diario (day 1 = the next business day; no Saturdays, Sundays or calendar holidays — the same count as the calculator with N = 15). The term is counted from the Estado Diario because that is how INAPI counts it: its Res. Ex. INAPI 138/2022 says the e-mail is a courtesy notice that does not replace the legal notification. For example, the partial acceptance of file 1554544, published on 2024-05-15, gives 2024-06-06 (the holiday of 2024-05-21 does not count), the same date INAPI wrote as the expiry in its resolution. It carries its three norms: Ley 19.039 art. 17 bis B (term: the appeal must be filed within fifteen days of the notification of the resolution), Ley 19.039 art. 11 (business_days) and Ley 19.880 art. 25 (computation). It is the SAME date as the appeal_close milestone in milestones[] (Milestones); this operation adds what a milestone cannot carry.

Example — file 1606887 (up to its appeal: the REC03 arrived on day 16 and was declared late):

{
  "applicationNumber": "1606887",
  "todayChile": "2026-09-29",
  "deadlines": [
    {
      "kind": "appeal_close",
      "date": "2025-09-16",
      "basis": "derived",
      "anchor": { "kind": "rejection_notice", "date": "2025-08-26", "annotationId": "8841468", "sectionCode": "024" },
      "rule": "15_business_days_after_notification",
      "legalBasis": [
        { "norm": "Ley 19.039", "article": "17 bis B", "role": "term" },
        { "norm": "Ley 19.039", "article": "11", "role": "business_days" },
        { "norm": "Ley 19.880", "article": "25", "role": "computation" }
      ],
      "businessDaysLeft": 0,
      "elapsed": true,
      "appealObserved": { "annotationId": "8841469", "date": "2025-09-17" },
      "caveat": null,
      "laterResolutions": []
    }
  ],
  "omitted": [],
  "calendar": { "table": "feriado", "coverage": { "from": "2013-01-01", "through": "2030-12-31" } },
  "meta": { "dataAsOf": "2026-09-29T11:59:09.379Z" }
}
  • anchor names the annotation of the anchoring resolution (annotationId) and says which one it is with sectionCode: "024" (rejection) or "444" (partial acceptance; the appeal is against the part of the coverage that was not accepted). anchor.kind is rejection_notice in both cases. Audit the date with GET /annotations/{id}.
  • When the 024 or the 444 is the MOST RECENT resolutive movement of the mark (the family 024, 444, M6, M11, 664B, 662B, 484; between a 024 and a 444 the most recent one anchors and, on the same date, the 024 wins). When there is neither to start from, there is no date and omitted[] says why, with one of four reasons: latest_resolution_not_anchoring (the latest is neither a rejection nor a partial acceptance), madrid_designation, anchor_on_non_business_day or outside_holiday_calendar — the last three apply to a 444 exactly as to a 024. That is «not known», never «no deadline»: the core does not estimate this date.
  • ⭐ Resolutions after the anchoring one («option C», 2026-09-29): when the latest is an M11 (e.g. «no ha lugar por extemporáneo»), 664B, 662B, 484 or an M6 without a 024 the same day, and an earlier 024 or 444 exists, the date is given, computed from the most recent of the two, with caveat: "later_resolution_on_record" and laterResolutions[] (annotationId, sectionCode, sectionName, date, oldest first): the date may no longer apply. A fact with its context, not a verdict. In that case there is no appeal_close milestone in milestones[]. Without the caveat, caveat is null and laterResolutions is [].
  • appealObserved is the first appeal (REC03) the corpus records after the anchor: a fact, not a verdict (it says neither whether it was timely nor whether it will succeed).
  • todayChile, elapsed and businessDaysLeft are evaluated today, Chile time.
  • The core applies no safety margin and notifies no one: whom to warn and how early is the consumer's call. An unknown mark is 404 not_found; without brands:read, 403 before any quota is used.

Payment of the grant fees

⭐ Since 2026-10-10 the same operation gives a second deadline, kind: "grant_fee_close": the last day to prove payment of the grant fees of an accepted application — the 60th business day STRICTLY after the notification of the acceptance resolution (029) or, after a partial acceptance (444) or an opposition ruling (484, 485), of the certificate of finality (911, 912). Unpaid, the application is held abandoned. Its norms: Ley 19.039 art. 18 bis E (term), Ley 19.039 art. 11 (business_days) and Ley 19.880 art. 25 (computation). It is the SAME date as the grant_fee_close milestone of milestones[].

The reservation, always in sight. The law counts the 60 days «desde la fecha en que quede ejecutoriada la resolución que autoriza la inscripción» (from the day the resolution becomes final); INAPI writes in its resolution that they are counted «desde la fecha de la notificación de esta resolución» (from its notification). The published date follows what INAPI applies: that is why it is always estimated (estimateReasons: ["anchor_by_office_practice"]) and carries reservation, with its code and its text in Spanish, English and Portuguese. ⛔ Show the date with that reservation, never as a certain date. For example, the acceptance of file 1633868, notified on 2026-03-10, gives 2026-06-05 — the day INAPI recorded as «Fin de plazo»:

{
  "kind": "grant_fee_close",
  "date": "2026-06-05",
  "basis": "estimated",
  "estimateReasons": ["anchor_by_office_practice"],
  "anchor": { "kind": "acceptance_notice", "date": "2026-03-10", "annotationId": "9886756", "sectionCode": "029" },
  "rule": "60_business_days_after_notification",
  "legalBasis": [
    { "norm": "Ley 19.039", "article": "18 bis E", "role": "term" },
    { "norm": "Ley 19.039", "article": "11", "role": "business_days" },
    { "norm": "Ley 19.880", "article": "25", "role": "computation" }
  ],
  "reservation": {
    "code": "statute_counts_from_finality",
    "text": { "es": "La ley cuenta los 60 días «desde la fecha en que quede ejecutoriada la resolución que autoriza la inscripción» […]", "en": "The law counts the 60 days […]", "pt": "A lei conta os 60 dias […]" }
  },
  "businessDaysLeft": 0,
  "elapsed": true,
  "paymentObserved": null
}
  • anchor names the act (annotationId, sectionCode: 029, 911 or 912); the most recent one rules. anchor.kind is acceptance_notice (029) or finality_certificate (911, 912).
  • paymentObserved is the first final-fee payment (PAG04) the corpus records from the anchor on: a fact, not a verdict (it does not say whether it was timely, and null does not say it was not paid).
  • With no date, omitted[] carries kind: "grant_fee_close" and one of five reasons: acceptance_not_final (the latest act is a partial acceptance or an opposition ruling and its certificate of finality is not in the corpus yet: the term runs from it), madrid_designation (the fees of a designation are paid through WIPO), prior_legal_regime (accepted before 2022-05-09: INAPI counted otherwise), anchor_on_non_business_day or outside_holiday_calendar.

Answer to the transfer of an opposition

⭐ And a third deadline, kind: "opposition_response_close": the last day for the applicant to answer the transfer of an opposition — the 30th business day STRICTLY after the transfer was notified, not extendable. Its norms: Ley 19.039 art. 9 (term), Ley 19.039 art. 11 (business_days), Ley 19.880 art. 25 (computation) and Decreto 82/2022 art. 24 (regulation). ⛔ Not answering is not abandonment: once the term has elapsed, «se haya o no evacuado el traslado», the procedure goes on. It is the SAME date as the opposition_response_close milestone of milestones[] (Milestones) and as the term of the facts of a filing. For example, the transfer of file 1605530, notified on 2025-05-22, gives 2025-07-04 — the day INAPI recorded as «Fin de plazo»:

{
  "kind": "opposition_response_close",
  "date": "2025-07-04",
  "basis": "derived",
  "estimateReasons": [],
  "anchor": { "kind": "transfer_notice", "date": "2025-05-22", "annotationId": "11199786", "sectionCode": "C1" },
  "rule": "30_business_days_after_notification",
  "legalBasis": [
    { "norm": "Ley 19.039", "article": "9", "role": "term" },
    { "norm": "Ley 19.039", "article": "11", "role": "business_days" },
    { "norm": "Ley 19.880", "article": "25", "role": "computation" },
    { "norm": "Decreto 82/2022", "article": "24", "role": "regulation" }
  ],
  "businessDaysLeft": 0,
  "elapsed": true,
  "answerObserved": { "annotationId": "…", "date": "2025-06-17" }
}
  • anchor names the act (annotationId) and its code: C1, the Estado Diario notification, when the corpus has it; otherwise the transfer resolution (458, 647, 1197, 1198). The most recent transfer of the mark rules.
  • basis is derived when the corpus has, on the same day, the notification and the resolution, and estimated when one of the two is missing; estimateReasons says which was inferred (notification_inferred_from_resolution_date or transfer_resolution_not_in_corpus) and is [] when it is derived. Show an estimated one as an estimate.
  • answerObserved is the first answer (OPO05) the corpus records from the anchor on: a fact, not a verdict (it does not say whether it was timely, and null does not say it was not answered).
  • With no date, omitted[] carries kind: "opposition_response_close" and one of four reasons: madrid_designation (INAPI sets another term for a Madrid designation), prior_legal_regime (a transfer before 2022-05-09, notified by registered letter), anchor_on_non_business_day or outside_holiday_calendar. A mark with no transfer in the corpus has no entry at all: the transfer reaches the corpus weeks after the opposition.

INAPI's official fees

POST /fee-schedule/calculate · scope brands:read · MCP equivalent: calculate_fees

Computes INAPI's official fees for a filing or a renewal over the table of chapter 13 of the Guidelines (Fees): the amounts are the table's literal UTM values and the pesos are derived at read time with the month's UTM, exact and not rounded.

curl -s -X POST "$TARNO_BASE_URL/fee-schedule/calculate" -H "X-API-Key: $TARNO_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"procedure": "application_and_registration", "niceClassCount": 3}'
# → lines: 1 UTM × 3 (application) + 2 UTM × 3 (registration); totalUtm 9; totalClp = 9 × the month's UTM

curl -s -X POST "$TARNO_BASE_URL/fee-schedule/calculate" -H "X-API-Key: $TARNO_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"procedure": "renewal", "applicationNumber": "1234567", "on": "2026-09-11"}'
  • procedure: application_and_registration (application + registration, per class) or renewal.
  • Exactly one of niceClassCount (1–45) or applicationNumber (its classes are counted).
  • on: the payment date (today in Chile by default). utmMonth: the UTM month (default: that of on); if not published yet, the latest earlier one is used and utm.basis says so.

Never included — always listed in excluded: the publication in the Diario Oficial (not an INAPI fee) and professional fees.

Renewal and the surcharge

With applicationNumber, a renewal uses the trademark's published expiry and the SAME rule as its renewalStatus (Ley 19.039 art. 24), evaluated on the payment date:

  • Within the term (window_open, or not_due before it opens): 6 UTM per class and the total.
  • In the grace period (grace): the art. 18 bis B (2nd para.) surcharge accrues «per month or fraction of a month after expiry». surchargeMonth is the payment month counted in calendar months from the expiry, date to date (one day after expiry is already month 1; never 30-day blocks). The row of that month of INAPI's Directrices 2026 surcharge table (chapter 13; chapter 23 §42) applies as the rule in force (renewal.surchargeRuleSource: "Directrices 2026", renewal.ruleIsOpenQuestion: false), and the total is computed. The table's «(*)» note and other INAPI fee publications word it differently: that discrepancy travels as an informative note in renewal.surchargeDoubt (with the literal notes) and the row carries subjectToDoubt: true (months 2 to 6; month 1 — 6 UTM per class, no surcharge — was confirmed by INAPI in writing on 2026-11-04 and carries false), but it does not block the total.
  • After the grace period (lapsed): lines is empty and the total null ("lapsed").
  • Renewal already granted, expiry not yet updated (renewed_expiry_pending, before or after the expiry): lines is empty and the total null ("renewed_expiry_pending").

With niceClassCount no expiry is known: the renewal is priced within the term and renewal.expiresOn is null.

Errors: unknown trademark ⇒ 404; a trademark the renewal rule does not apply to (not registered, Madrid route, no expiry) ⇒ 422 renewal_not_applicable; no classes ⇒ 422 no_nice_classes; both or neither of niceClassCount/applicationNumber ⇒ 400.

Outcomes of a party's trademarks

GET /parties/{id}/outcomes · scope brands:read · MCP equivalent: get_party_outcomes

«How do the trademarks this representative files end up?», counted from the corpus with the definition in view. Each trademark in which the party appears in role role (representative by default, or holder) and relation relation (all by default — the track record includes retired relations; current, former) counts once, by its published status today (the same as status), narrowed by niceClass and filedFrom/filedTo (echoed in population.filters).

group published statuses
registered_ever concedida, registrada, esperando_renovacion, vencida, caducado
refused rechazada, denegada
withdrawn_or_abandoned desistida, abandonada
annulled anulada
pending en_tramite, observacion_de_fondo, publicada, oposicion
  • decided = registered_ever + refused + withdrawn_or_abandoned + annulled; pending trademarks and those without a status (unclassified) are outside the denominator.
  • grantShare = registered_ever / decided; refusedShare, withdrawnShare and annulledShare over the same denominator. With decided = 0 all four are null, never 0.
  • definition carries the full definition in three languages, to be quoted.

⛔ It is not a «success rate» nor a judgement on the party: a trademark's outcome depends on many factors beyond who files it. Always quote population.trademarks (the n).

Statistics of a class

GET /nice-classes/{niceClass}/stats · scope brands:read · MCP equivalent: get_class_stats

"In class 25, of the applications filed in 2021–2025, what share was registered, how long did the median take, and how many had an opposition?" — computed from the corpus, with its n, its date and what it does not cover.

curl -sS "$TARNO_BASE_URL/nice-classes/25/stats?filedFromYear=2021&filedToYear=2025" \
  -H "X-API-Key: $TARNO_API_KEY"
  • The period is in whole filing years (filedFromYear, filedToYear). By default, the last five complete years in Chile (today, 2021–2025), returned in period with defaulted. filedFromYear > filedToYear ⇒ 400 invalid_range; a class outside 1–45 ⇒ 400 validation_error.
  • The population (population.filed, the n): applications of category marca with the class filed in the period. The Madrid Protocol and the other categories are left out; those without a category yet are counted apart (excludedUnknownCategory).
  • Outcomes: the same groups as the outcomes of a party (same list of published statuses, same decided, grantShare and other shares, null when nothing is decided).
  • Time to registration (timeToRegistration): percentiles p25/p50/p75 in days of the trademarks with a registration date, with their n and the histogram they come from. ⚠ They are approximate (approximate: true): they come from 30-day buckets, with a maximum error of 15 days (precisionDays). Never a mean. Registration dates before the filing date are left out and counted (excludedInvalidDates).
  • Acts (actsShare): the share of the population with at least one substantive objection (substantiveObjection), formal objection (formalObjection) and opposition (opposition) annotation. A trademark with several acts of a family counts once. Opposition is counted by annotations, never by relations (which only exist since 2024-04-19), and always carries its coverageNote.

Coverage, stated — coverage

  • annotationSources: the sources the acts come from. There are two: INAPI's record pages (each trademark's full history, buscador) and the Estado Diario (since 2024-04-19, estado_diario).
  • windows: the date windows in which the acts come from one source only. Today it is an empty list: there is none.
  • notes: what this answer does not cover, with a stable code and its text in three languages:
code when
madrid_and_other_categories_excluded always
unknown_category_excluded some filings have no category yet (the most recent ones, and all before 2009)
current_year_partial the period includes the current year
substantive_objection_history_limited the period starts before 2012: until February 2012 INAPI did not publish the substantive objection as an act of its own. For those years the figure counts, besides the trademarks with substantive-examination acts, the trademarks whose final decision (refusal W160 or partial acceptance W152) cites a ground of articles 19 or 20 of Law 19.039. It is a lower bound, because an objection that was overcome left no act (measured on 2026-10-10, all classes: 3,599 of 52,975 of 2009, 4,341 of 57,501 of 2010, 6,940 of 67,859 of 2011). The share is published as it is, never filled in
refreshed_before_latest_sync a corpus sync started after the computation

Where it comes from and when — dataAsOf

The figures are computed once per sync, at the end of the chain, in a view of the corpus (aggregating them per request cost 2.5 s for class 35, measured in production). dataAsOf is the time of that computation. Before the first computation the operation answers 503 class_stats_pending: never half-made figures.

⛔ It is a descriptive signal of the class, not a prediction for a given application. Always quote population.filed and the period.

Compare two trademarks

POST /trademarks/compare · scope brands:read · MCP twin: compare_trademarks

Two trademarks, channel by channel, with the SAME measures and folding (accents and case ignored) as the similarity engine. Each side (a, b) is exactly one of:

  • { "text": "…", "niceClass": [35] } — a text (1–200 characters, at least one letter or digit) with optional classes;
  • { "applicationNumber": "1510102" } — a trademark of the corpus: its denomination, classes, goods and services and stored semantic vector. niceClass with a number is a 400; an unknown number, 404.
curl -s -X POST "$TARNO_BASE_URL/trademarks/compare" -H "X-API-Key: $TARNO_API_KEY" \
  -H "content-type: application/json" \
  -d '{"a":{"text":"ALTODESEMPEÑO CONSULTORES","niceClass":[35]},"b":{"applicationNumber":"1510102"}}'

The channels — signals

channel what it measures
lexical.wordSimilarityAB / …BA word trigram similarity both ways (it is not symmetric): how much of A appears as a stretch of B, and the reverse
lexical.trigramSimilarity trigram similarity over the whole texts
editDistance.levenshtein / normalized edits to turn A into B (Levenshtein distance) and the same figure over the longer text's length (0 = identical)
phonetic.wordSimilarity / keysEqual the Spanish phonetic key (seseo, yeísmo, betacismo, silent h, qu/k): keysEqual: true = they sound the same
semantic semantic similarity (cosine) rescaled against the unrelated-pairs baseline (method.semanticBaseline, 0.78)
sharedTokens the folded words present in both

classes gives the Nice classes in common and apart (null when a side has none). With two trademarks by number, goodsAndServices.perSharedClass lists, per shared class, the Spanish word stems present in both descriptions («camisas» gives camis), capped at 50 (truncated).

Semantic. A text goes as query: and the other as passage: (the baseline was calibrated that way); a trademark by number brings its stored vector without calling the sidecar. meta.semanticPairing states the pairing (query_passage, or passage_passage for two trademarks). ⚠ semantic: null means not evaluated, never a 0: meta.channels.semantic says why — degraded (the sidecar took more than 800 ms or failed; the rest is served), not_configured (no sidecar) or not_embedded (the trademark has no vector yet).

What it does NOT return, and why

No global score, no level (high/medium/low), no «risk» and no article of law as a ground. Each channel is a signal measured with its function; turning them into a confusion verdict is a legal decision the corpus does not take. method explains every channel in three languages, to be quoted.

Generic terms and the distinctive core

Each side declares which words do not distinguish in ITS classes, and why (genericTerms[].reasons):

  • curated — the neutral list approved by the Tarno team (method.genericTerms.listVersion): connectors (de, la, the…), company forms (ltda, grupo… and the phrases «s.a.», «s.p.a.», «e.i.r.l.» recognised before tokenizing), demonyms (chile…), domains (cl, com); commerce (store, tienda, distribuidora…) only in 35 and 25; professional services (consultores, servicios, soluciones…) only in 35 and 42; spa only as the last word and never in classes 3 or 44. Never mas, plus, pro nor single letters.
  • corpus_frequency — sector vocabulary: the word is in at least max(50, 0.75 % of the class's trademarks) (count, classSize, threshold), counted on the engine's index with a cap. So clinica in 44 or hotel in 43 are discounted without being listed.

decompounded splits a joined word of 8 or more letters only if both parts exist as words of the corpus and the whole word does not («altodesempeno» ⇒ «alto» + «desempeno»). distinctiveCore is what is left (coreStatus: reduced); null when there was nothing to discount (unchanged) or only generic words were left (only_generic_terms) — never an empty core presented as distinctive.

distinctiveSignals repeats the lexical, edit and phonetic channels over the cores (no semantic). It is a declared discount: signals never change.

The case behind it. The availability report of «ALTODESEMPEÑO CONSULTORES» in class 35 returned 40 of 40 similar marks because of «CONSULTORES» and missed «MAS Desempeño» (measured in production on 2026-09-28). Compared, A declares consultores generic in 35 and altodesempeno split; its core is «alto desempeno», and distinctiveSignals shows what they share: sharedTokens: ["desempeno"].

Scope: brands:read. One call = one quota unit.

Recommend Nice classes

POST /nice-classes/recommend · scope brands:read · MCP equivalent: recommend_nice_classes

"I am going to sell sportswear online: in which classes?". A free-text description (3 to 500 characters, Spanish) ⇒ the Nice classes of the edition in force, each with the official WIPO term, the heading and real examples from the corpus, with each route separately.

curl -sS -X POST "$TARNO_BASE_URL/nice-classes/recommend" \
  -H "X-API-Key: $TARNO_API_KEY" -H "Content-Type: application/json" \
  -d '{"description": "venta de ropa deportiva por internet", "limit": 5}'

The edition (edition, editionSelection): the one in force in Chile today. It is 13-2026 until 2026-12-31 and 13-2027 from 2027-01-01, with nobody changing anything. The answer says which one it used and since when it applies. ⛔ Never the 12th edition nor INAPI's classifier of earlier editions. listVersion is the version of the loaded official list (20260101), or null when that edition is not loaded yet. The headings (heading) are those of that same edition: from 2027-01-01, the 13-2027 ones, with pt-BR (Tarno's own translation) there too.

The four routes, each in signals and its position in ranks:

route what it compares value
termSemantic the description with the official WIPO terms (≈ 10,000, with their basic number), by meaning raw semantic similarity (cosine) of the class's closest term
termLexical the words of the description with the text of those terms the best text rank (0–1) of a term of the class
headingSemantic the description with the 45 headings, by meaning raw semantic similarity (cosine)
corpusEvidence the goods and services descriptions of real trademarks of the corpus that match the description (the first 1,000, method.corpusSample) how many are of the class, over how many were looked at

Each route orders the classes and they are fused with RRF (k = 60, the same rule as the search). For each candidate class: officialTerms[] (up to 5: basic number, term in es and en, and the route that found it) and corpusExamples[] (up to 3: application number and a literal excerpt of its description, with excerptOffset). ⚠ The semantic cosine compresses unrelated pairs into ~0.75–0.80: use it to compare classes with each other, not as an absolute score.

When something is missing, it is said (meta.channels):

  • degraded — the vector service failed or took more than 800 ms: the meaning routes are null and the others are served anyway.
  • pending_embeddings — the official list is still being prepared for search by meaning (at night, between 00:00 and 07:00 in Chile, a few days after each load). meta.pendingEmbeddings says how many rows are missing. The list's text route goes on.
  • not_loaded — no list is loaded for the edition in force: only the corpus evidence runs.
  • not_configured — the deployment has no vector service.

Attribution (attribution): the terms and the headings are WIPO's (Nice Classification, CC BY 4.0). Quote the attribution when you show them. dataAsOf.officialList is the list's date; dataAsOf.corpus, the corpus's.

Edges: a description with only stop-words (de la para) or only exclusions (-calzado) ⇒ 400 validation_error; fewer than 3 or more than 500 characters, or limit outside 1–10 ⇒ 400.

⛔ It is a signal to choose classes, not an official classification: the applicant decides it and INAPI reviews it.

Permission: brands:read. One call = one quota unit.

POST /logos/analyze · scope logo:analyze · MCP equivalent: analyze_logo

⛔ The description is generated by a language model (Anthropic, claude-sonnet-5 by default), not by INAPI nor by a person. It is the only operation of the contract that uses a language model, and the response says so in generatedBy and disclaimer. Quote the disclaimer with the result.

An image ⇒ what it shows, in INAPI's terms: the word elements (the text, literally), the figurative elements, the colours, the lettering and the sign type the model suggests (denominativa, figurativa, mixta, etiqueta, tridimensional or null); and the codes of the Vienna Classification, 10th edition (VCL10), in force since 2026-01-01.

The style of description (since 2026-09-29) is that of the label descriptions in INAPI's register, measured on 99,996 descriptions of mixed and figurative marks filed in 2021–2024: it opens with «Etiqueta consistente en …» (the most frequent opening: 40.6 % of mixed and 50.5 % of figurative marks; «La marca tridimensional consiste en …» for a three-dimensional shape, and one short sentence starting «Denominación "…"» for a words-only sign, which carries no description in the register); it walks the elements in reading order and places them relative to each other («sobre», «bajo», «en su interior», «a la izquierda», «en la parte superior»); it quotes the text literally; it describes the lettering in plain words («letras mayúsculas de color negro», «letras estilizadas»); it attaches each colour to its element, with Chilean names («naranjo», «celeste», «café») and no Pantone/RGB codes; and it closes with the background («Todo sobre fondo blanco.»). One to three sentences, about 150–300 characters (the register's median is 212). ⛔ It does not write the formula «se reivindican los colores» (0.1 % of the register) nor statements about the scope of protection («sin protección a las demás palabras…»): they do not describe the image. The text is in Spanish, as in the register.

curl -sS -X POST "$TARNO_BASE_URL/logos/analyze" \
  -H "X-API-Key: $TARNO_API_KEY" -H "Content-Type: application/json" \
  -d "{\"image\": {\"mediaType\": \"image/png\", \"dataBase64\": \"$(base64 < logo.png | tr -d '\n')\"}}"

The image goes in the body, base64 (no data: prefix, no line breaks): PNG, JPEG, GIF or WebP, checked by the file's header (a PNG declared as image/jpeg is a 400), up to 700 KB decoded and 2,000 px per side. ⛔ Never a URL: the core does not download images.

What happens to the image: it is sent to Anthropic for the analysis (together with the Vienna catalogue and the instructions; nothing else: neither your key nor who is calling) and it is not stored nor logged anywhere in the core. Only its fingerprint comes back: image {mediaType, width, height, bytes, sha256} — use the sha256 to avoid asking twice for the same logo.

The Vienna codes (viennaCodes): only those the model proposes that exist in VCL10, each with WIPO's label in its two authentic languages (en, fr), never with the model's words. ⚠ es is always null: neither WIPO nor INAPI publishes Spanish labels, and the core does not translate them. discardedCodes counts the proposed codes that do not exist (dropped, not shown). viennaCatalog.attribution is WIPO's copyright acknowledgement: quote it with the codes.

Own quotas (on top of the unit of the monthly quota, like any call): 20 analyses per key per Chilean day, and 200 per day across all keys. Going over ⇒ 429 logo_quota_exceeded (the key) or 429 logo_global_quota_exceeded (the service), with Retry-After until midnight in Chile and retryAfterSeconds in the body — different from the monthly 429 quota_exceeded.

Edges:

status code when
400 invalid_image the image is not canonical base64, not PNG/JPEG/GIF/WebP, does not match mediaType, exceeds 700 KB or 2,000 px (reason says which)
400 validation_error the input has an extra field (a url, e.g.) or a mediaType outside the four
422 logo_analysis_refused the model declined to describe the image
422 logo_image_rejected the provider could not process the image (it passed the header check, but the file is damaged or not what it claims): do not retry the same image. The call was made and counts against the daily quota
429 logo_quota_exceeded / logo_global_quota_exceeded the daily quota
502 logo_analysis_invalid_output the model's output did not match the schema (always validated; free text is never interpreted)
503 logo_analysis_not_configured the server has no provider key: the operation exists and answers this
503 logo_analysis_unavailable the provider does not answer or is overloaded: retry later
503 logo_quota_unavailable the quota counter does not answer: without a counter the model is not called

⛔ It is not an INAPI decision and gives no opinion on whether the mark can be registered, is distinctive or conflicts with another one. It describes the image and proposes codes; INAPI classifies.

Permission: logo:analyze, granted explicitly (an unrestricted key does not have it, nor the trial key, nor reports:read). One call = one monthly quota unit + one of the day.

Resemblance by image (future operation)

POST /logos/similar · scope logo:similar · MCP equivalent: find_similar_logos

⛔ FUTURE OPERATION, not available yet. "Give me the trademarks whose logo looks like this image" is declared in the contract —its input and the shape of its output— so that you can program against it, but today it has no procedure: with a valid image and the scope it always answers 503 image_similarity_not_available, never results. When it exists, a new entry of the contract changelog will announce it.

curl -sS -X POST "$TARNO_BASE_URL/logos/similar" \
  -H "X-API-Key: $TARNO_API_KEY" -H "Content-Type: application/json" \
  -d "{\"image\": {\"mediaType\": \"image/png\", \"dataBase64\": \"$(base64 < logo.png | tr -d '\n')\"}, \"niceClasses\": [25], \"limit\": 20}"
{
  "type": "https://api.tarno.cl/problems/image_similarity_not_available",
  "title": "Service unavailable",
  "status": 503,
  "detail": "Image similarity search is a declared future operation and is not available on this server yet.",
  "code": "image_similarity_not_available",
  "requestId": "…"
}

The input: image {mediaType, dataBase64} —the same as in analyze_logo: PNG, JPEG, GIF or WebP in base64, checked by the header, at most 700 KB and 2,000 px per side, never a URL—; niceClasses (optional, Nice classes 1..45: only trademarks with any of them) and limit (optional, 1..50). The image is validated by the same function as analyze_logo: a bad image is the same 400 invalid_image with the same reason. It is not stored nor logged.

The future output (not emitted today), in the usual collection envelope:

  • results[]: applicationNumber, denomination, niceClasses, signType, imageUrl, score and signals.visual (between 0 and 1) and matchedBy: ["visual"]. The score is a visual resemblance: a signal, never an opinion on whether the mark can be registered or conflicts with another one.
  • meta: dataAsOf, model {name, dimensions, embeddedAt}, coverage {trademarksWithImage, trademarksEmbedded, asOf} (what part of the corpus the search could see: a trademark outside it is not "no resemblance"), disclaimer (three languages) and image {mediaType, width, height, bytes, sha256}.

Edges today:

status code when
400 invalid_image the same image validation as analyze_logo (reason says which)
400 validation_error an extra field (a url, e.g.), classes outside 1..45 or limit outside 1..50
403 forbidden the key does not have logo:similar
503 image_similarity_not_available always, with a valid image and the scope. ⛔ Do not retry: unlike service_unavailable, it is not a passing outage

Difference with analyze_logo: analyze_logo describes an image with a third-party model and proposes Vienna codes; find_similar_logos will search the corpus for trademarks with a similar logo. They are separate scopes: logo:analyze does not open logo:similar, nor the other way round.

Permission: logo:similar, granted explicitly (not an unrestricted key, nor the trial key, nor reports:read, nor logo:analyze). ⚠ It is granted to no key while the operation is not available: do not ask for it before its announcement.