Citation check: does the cited source support the claim?

Decision trace

What this lab demonstrates

A citation is useful only if a reader can resolve it and the source actually supports the sentence. This teaching fixture uses a claim that a cache expires after sixty seconds and a source that states that value. It also supplies four deliberate defects: a malformed identifier, a reference to a missing source, a six-hundred-second claim attached to the sixty-second source, and an outdated source version. Each stage asks a different question. A link that opens is not enough to establish factual support, and a matching sentence from an old version may be unsuitable for a current operational instruction.

How to use each control

Injected defect selects one of the fixed citation failures or the clean baseline. Checks included decides how deep the reviewer goes: one checks identifier form, two also resolve the source, three compare the number in the claim, and four examine source freshness. Play test run advances through the checks in a sixty-second visual sequence. Pause holds the view, Restart begins again, and the Timeline slider seeks to a stage directly. Speed changes only viewing pace. The decision trace lists each check and whether it ran, passed, or caught the injected defect.

Try a controlled failure

Start with the six-hundred-second claim and include only two checks. The reference syntax and source lookup pass, but the numerical mismatch escapes because no claim comparison ran. Include the third check and replay: the mismatch is caught. Then select Old source version cited. Even a full claim comparison cannot see the version problem until the fourth check is included. Finally test a missing source with depth one and then depth two. Record which exact observation changes the result; do not conclude that more green checks always mean a better-supported answer.

What this model simplifies

The source text, claim, identifier, and version are fixed teaching values. This page does not retrieve a website, parse a PDF, inspect a real ADK Java answer, or decide semantic entailment with a model. A production citation checker needs stable source identifiers, access to the cited span, version and publication context, and a policy for paraphrase, numbers, and conflicting evidence. Even a supported quotation can be used misleadingly if crucial qualifiers are omitted. The sixty-second sequence is for inspection, not measured verification latency or a guarantee that the entire answer is correct.

A formatted reference can still support the wrong claim

Citation verification is split into reference syntax, source existence, claim support and source version. A claim that cache expiry is six hundred seconds cannot be supported by the fixture source that states sixty seconds. A resolvable identifier is therefore weaker evidence than a supported claim. These stages illustrate four independent review obligations using predetermined defects. The browser does not fetch cited pages, parse an arbitrary answer or run a semantic entailment model.

Choose coverage deliberately

Injected defect selects one known fault, or a clean fixture. Checks included runs the first one, two, three or all four checks in their displayed order. This is prefix coverage, not a random sample and not a selection of arbitrary individual checks. A skipped check provides no evidence. The model uses an explicit mapping from defect identifiers to checks that detect them; it does not execute Java code, call an LLM, compile a package or infer semantic correctness. An escaped defect is visible because the simulator knows which fault you injected, whereas a real test runner would need independent evidence to discover that escape.

Reproduce a coverage gap

Select the 600-second claim citing the 60-second source and include two checks. Syntax and source resolution pass, but claim comparison is skipped and the defect escapes. Include three checks to catch the mismatch. Now choose Old source version cited at coverage three: support still passes in this fixture, but the missing freshness check permits an escape. Include all four to block it. A malformed identifier is caught immediately at the first selected check.

Interpret the failure boundary

Freshness alone cannot repair a mismatched numerical claim, and existence alone cannot establish support. Try every fault with full coverage and read which check fails. This fixture assumes a perfect detector for each named problem and a single source version; it cannot evaluate paraphrases, conflicting sources, quotation context or nuanced qualifications. A production citation checker must retain the actual evidence and claim text. The displayed outcomes explain a gate design rather than providing a measured hallucination or citation accuracy benchmark.

Inspect the replay

Play starts the sixty-second inspection and becomes Pause while running. Pause holds the current stage; Restart returns to the beginning without changing the experiment settings. Speed changes viewing pace rather than model outcomes. Timeline seeks directly to a stage and pauses playback. Changing a scenario setting recomputes the deterministic fixture and rewinds it. The final decision appears at the end, so move Timeline to sixty to compare completed runs. Stage cards and the text trace describe the same selected fixture; their elapsed viewing time is not measured agent execution latency.