Citation validation: resolve a source span, then test what it supports

Store a logical source identity, immutable version or content digest, source location, parser lineage, and span coordinates. Name whether offsets refer to bytes, Unicode code points, parsed-text characters, or tokenizer positions. A quotation should be checked against the exact referenced version using the declared mapping. If a live URL changes, a stored version or authorized snapshot can preserve auditability. Do not silently resolve an old span against new text and present a coincidentally matching phrase as the original evidence.

Generated summaries and extracted snippets need their own lineage back to source spans. A summary identifier can be useful in the trace, but it is not an original quotation. Citation resolution must enforce the current principal's access and retention policy, including historical snapshots. A link that opens for an administrator may fail or expose too much for an ordinary reader. Keep integrity and permission checks separate: a matching content hash establishes an identity condition, not that the viewer is entitled to the content.

Split an answer into factual claims at a useful granularity. For each, inspect whether the cited evidence entails the claim with its qualifications, units, time, and scope. A passage mentioning the same entity may not support the stated relationship. A quotation can be accurate while its surrounding context negates or limits the answer's interpretation. Multiple passages may jointly support a comparison or inference, so record that relation rather than demanding one sentence contain every fact. Unsupported does not always mean false; it means the designated evidence is insufficient.

Authority is another dimension. An internal draft, a public announcement, and an authoritative current record can disagree while all citations resolve correctly. Evaluate source relevance and precedence under the task's stated policy. ALCE studies citation quality and answer generation with explicit evaluation dimensions, while attribution work distinguishes support by a source from other aspects of correctness. Automated entailment checks are useful measured tools with their own errors. Calibrate them against human-labeled cases instead of treating a high verifier score as a proof.

Test nonexistent source IDs, mismatched versions, shifted spans, fabricated quotations, citation reuse across unrelated claims, and claims combining incompatible dates. Check that rendering retains the relation between claim and citation after editing or summarization. Verify the reader can access exactly the intended authorized artifact. Report mechanical resolution, semantic support, claim coverage, and source quality separately. A pipeline can improve link validity without improving factual grounding, or improve answer correctness while leaving important claims uncited.

The probability lab shows a hypothetical independent-check calculation only. Reset, set p=0.80 and n=10, predict expected eight successes and probability all ten succeed about 0.1074, then Play or Step. Separately Reset, set p=0.90 and n=10, predict mean nine and all-success probability about 0.3487, then Play or Step. Real citation checks can share correlated extraction and verifier errors. The lab has no source resolver or entailment model; use it to distinguish average success from a whole-answer condition under explicitly stated assumptions.