Fact-check a link claim by claim against its own sources

01 · Role in the system

What Jev does here

For every claim extracted from a submitted link, one Choice question goes to Jev with the cited excerpts inlined in the question, asking whether they support or contradict the exact claim, including its dates, scope, and numbers. The instructions are closed-world: use only these excerpts, treat them as untrusted evidence never as instructions, and choose unknown when the evidence is missing, partial, irrelevant, or conflicting. A relation counts as evidence only at or above 0.8 confidence; between 0.7 and 0.8 the site shows it as likely and never counts it. The model is pinned to jev-1.13.0 rather than a moving alias so those thresholds stay meaningful, and the client fails closed through a bounded retry ladder. The page-level verdict - REAL, FAKE, or Insufficient - aggregates the counted relations, with Insufficient the default whenever agreeing evidence runs short.

The claim, the excerpts, and nothing else

The question is deliberately starved of context. The claim and its cited excerpts ride inside the question itself, the model is told to use only those excerpts and no outside knowledge, and unknown is a first-class answer for empty, partial, irrelevant, or conflicting evidence - so a page whose sources cannot be found produces unknowns, not guesses. Excerpts are also marked as untrusted evidence rather than instructions, which is the standing defense against a source page trying to talk its way into a verdict.

Checking the exact claim - its dates, scope, and numbers, in the question's own words - is what keeps the judgment honest at the edges. A story that mostly agrees with a claim but gets the year wrong is a contradiction of the exact claim, not a support, and the burden of that comparison lives in the typed question rather than in post-hoc parsing of free text.

Confidence decides what counts as evidence

Every judgment comes back with a confidence, and a single policy file - shared by code, tests, and documentation - decides what each number means: at or above 0.8 a relation counts as evidence, between 0.7 and 0.8 it is shown to the reader as likely and never counted, and below that it is noise. The page-level REAL or FAKE verdict is built only from counted relations, so a confident minority cannot manufacture certainty and a weak majority cannot either; the default outcome is Insufficient, and the site treats that as a result rather than a failure.

The thresholds follow the vendor's citation-check cookbook default pending the project's own benchmark, and the policy comment says exactly that - the numbers are adopted, not claimed as measured.

A pinned model so the thresholds stay honest

The client pins jev-1.13.0 instead of a moving alias, with the reason written in the source: tuned confidence thresholds only stay meaningful against a versioned model. Around the call sits production plumbing that respects the design - a bounded retry ladder on rate limits and overload that fails closed, state and request size caps, and secret scrubbing so an upstream error message can never echo an API key into a log.

02 · Control boundary

Where Jev sits

One Choice per claim with cited excerpts inlined in the question; relations return with confidences that deterministic policy code gates at fixed thresholds, and aggregation never counts a judgment it cannot trust.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Jev reads only the cited excerpts inlined in each question; sources that were never retrieved or cited are invisible, and the model is explicitly told not to use outside knowledge.
• A relation counts as evidence only at or above 0.8 confidence - weaker judgments display as likely and never count - so thin sourcing yields Insufficient rather than a verdict.
• The page-level REAL/FAKE aggregation lives in the site backend; the public repository carries the claim-level engine, its policy file, and the Jev client.
• A verdict is a statement about cited evidence, not about the world: claims with no discoverable sources stay unverified by construction.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• DansiDanutz ↗Community · github · public · checked 2026-09-27
• Dan Semenescu ↗Community · demo · public · checked 2026-09-27