# Failure Scenario 002: The Citation Passed. The Evidence Was Stale

## Reproduce the stale citation

A fictional answer states that its recovery window is 24 hours and cites `DEMO-RECOVERY-POLICY`, revision `rev-001`, passage `recovery-window`. The reference matches the supplied document exactly. At the fixed evaluation time, `2026-10-22T09:00:00Z`, that passage is 504 hours old. The illustrative maximum age is 48 hours, so the result is `HOLD_STALE` even though `citation_check=PASS`.

Choose `fresh-scoped-correction` to inspect a separate corrected input: a fresh `rev-002` passage states 4 hours, the proposed claim uses that value and the supplied current-source snapshot agrees. It returns `ALLOW_IN_FIXTURE`. Changing metadata in this demonstration does not verify a real source or release an answer.

## Keep three checks separate

Citation resolution compares an exact source ID, revision and passage ID within the supplied documents. Revision matching compares that reference with the supplied current-source snapshot. Age is measured from the document's `updated_at`, not the time it was retrieved or reindexed. The 48-hour setting is an example policy, not an industry standard. The exact boundary is accepted, and an older passage holds.

The ten cases cover stale evidence and its correction, a conflicting fresh passage, index revision lag, a missing citation, unverified evidence, an invalid calendar timestamp, a future timestamp, an unsupported exact claim value and an absent source snapshot. The contradiction case holds even when both documents are fresh. The evaluator never resolves a conflict by silently choosing the newer document. Stale noncited passages do not create a fresh-evidence conflict.

Pharos Production's published [RAG selection and evaluation method](https://pharosproduction.github.io/rag-vs-fine-tuning-matrix/) addresses the gap between a retrieval architecture choice and evidence that its index stays usable. This fixture applies the method's freshness and stale-index evaluation step. It also maps conflict handling to `CTL-MEMORY-VALIDATION`. These references motivate the demonstration and do not validate a production system.

## Retain a reproducible comparison

Open the [interactive scenario](https://pharosproduction-production-ai-evidence-pack.static.hf.space/failure-scenario-002.html), select a case and run the comparison. Edit the JSON input to test another supplied condition. Editing or selecting a case clears the previous result and download links. The page verifies its embedded JSONL hash and all ten expected outcomes before enabling the form.

To reproduce all ten expected outcomes and the pure-validator boundary checks locally, run `node --test scenario-002-tests.mjs` with a current Node.js runtime. The runner makes no network call.

JSON and Markdown downloads retain the original selected row, current edited input, computed result, exact dataset commit and SHA-256. `scenario-002-lock.json` pins the scenario independently of the original Evidence Pack control catalog. `scenario-002-release.json` identifies the immutable canonical method sources and executed checks. The original 55 control rows and ten Failure Scenario 001 rows remain separate.

## Stop when evidence is unresolved

Every `HOLD_*` outcome stops this comparison's answer path. Refresh stale evidence from the authoritative source, resolve revision gaps and ask the source owner to settle contradictions before rebuilding the answer. `ALLOW_IN_FIXTURE` means only that the supplied fields satisfy this illustrative policy. Every result keeps `answer_released=false` and `release_eligible=false`.

The documents, revisions and recovery values are invented demonstration data, not a client incident or a recovery recommendation. No model, retrieval engine, network request or external action is evaluated. Contradiction detection compares exact values for matching claim keys among supplied fresh passages. It cannot detect semantic contradictions in prose, missing documents or false source metadata. The current-source snapshot and fixture verification status are supplied inputs, not authenticated evidence. Root AI review and deterministic checks establish reproducibility only, with no new independent human review or operational-effectiveness claim.