Tanilo

Docs · Experimental

Assessing a claim with /evaluate

POST /evaluate gives a model-based assessment of a claim before your workflow acts on it. It is live in the free public beta at https://api.tanilo.io: rate-limited, no key, no account.

Read this first

Assessments are model judgments and can be wrong. They are not suitable for automatically authorizing actions. Sources listed in a response don’t feed the verdict, and a valid signature doesn’t establish that the assessment is correct.

Request

Make a request

Free during the public beta, rate-limited, no key. Each request runs model checks on Tanilo’s side; the request below is an example and was not sent while writing this page.

curl -sS https://api.tanilo.io/evaluate \
    -H "content-type: application/json" \
    -d '{ "content": "The Eiffel Tower is in Paris.", "min_confidence": 0.8 }'
  • content (required): the claim or text to assess, a string or an object, at most 4,000 UTF-16 code units. Longer input is refused with 413 input_too_long; nothing is cut.
  • min_confidence (optional, 0.0–1.0, default 0.8): caller policy only. It can lower act to verify; it can never produce act when the published gate is halt.
  • url is disabled: a request that includes it is refused with 422 url_input_disabled before anything is fetched. source is accepted for compatibility and ignored.
  • In the response, evaluation.recommendation carries the result, evaluation.policy records the gate inputs and the published mapping, evaluation.sources lists the checker rows, and the top-level signing block says whether a receipt was issued.
  • GET https://api.tanilo.io/evaluate returns the current request and error reference, free and without a model call.

Results

What it returns

  • A recommendation of act, verify or reject. Since 19 October 2026 the recommendation is derived from the same gate decision used in newly issued signed receipts.
  • An additional caller-confidence threshold can change act to verify, but cannot override a halt.
  • Confidence values are heuristic scores, not calibrated probabilities.
  • When signing succeeds, a signed receipt. The body policy is versioned separately; historical receipts keep their original gate values.

Inputs and failures

What it refuses

Since the 2026-10-28 release (see the changelog):

  • Input longer than 4,000 UTF-16 code units is refused with 413 input_too_long, not cut.
  • URL input is disabled: a request containing the url field is refused with 422, including when content is also supplied.
  • Without usage accounting the endpoint answers 503 and calls no provider.
  • act requires a usable answer from every claim-producing checker; otherwise the result is halt.

Keep separate

What it is not

Not a deterministic check

For a specific condition, such as a price against supplied terms, use /v1/verify-facts: no model, and checks over the inputs you send give the same result for the same inputs (registry lookups depend on the registry at the time).

Not the contract-term prototype

Checking what an extraction left out is a separate prototype. It is not available through /evaluate.

Not a current accuracy figure

The accuracy figure on the benchmarks page was measured on 28 May 2026 on an earlier pipeline and does not describe the current service.

Incident record

In August 2026 receipts were signed for evaluations that had not run. The full record and what changed.

Verifying a receipt

Check that a signed result hasn’t changed since it was signed.

Changelog

Dated releases, including the 2026-10-28 input and failure-handling changes.