### BLOCK 0: INPUT STATUS Mode: LIVE RUN. Query, bad answer, candidates, cutoff, and expected behavior are captured. The live scoring formula and corpus contents are not supplied. ### BLOCK 1: NORMALIZED INCIDENT - Query scope: enterprise customers — OBSERVED. - Bad claim: enterprise refunds are available within 14 days — OBSERVED. - Supporting source for that wording: consumer policy limited to individual subscriptions — OBSERVED. - Correct enterprise source: Candidate 7, below the first-four generation cutoff — OBSERVED. ### BLOCK 2: SENTENCE-LEVEL SOURCE TRACE | Claim | Supporting chunk | Trace status | Contradiction | Evidence state | |---|---|---|---|---| | Enterprise customers may request a refund within 14 days | Consumer Refund Policy | SUPPORTED for consumer scope; invalidly transferred | Enterprise MSA Guide says terms vary by contract | OBSERVED | ### BLOCK 3: POOL vs ORDERING VERDICT Primary stage: ORDERING_LOSS with WRONG_CHUNK_SELECTED. The correct enterprise source entered the recorded candidates but remained below the generation cutoff. This is not a pool miss. The consumer chunk then won despite a scope mismatch. ### BLOCK 4: ATOMIC CAUSE MAP | Cause | Failure class | Evidence signal | Evidence state | Explained content | |---|---|---|---|---| | Enterprise source ranked below cutoff | ORDERING_LOSS | Candidate 7 vs first-four cutoff | OBSERVED | correct rule absent from generation context | | Consumer source outranked scope-matched source | WRONG_CHUNK_SELECTED | Candidate 1 is individual-only | OBSERVED | 14-day claim transferred to enterprise | ### BLOCK 5: COUNTERFACTUAL RESULTS | Cause | Mechanism test | Removal test | Competing cause | Result | |---|---|---|---|---| | Enterprise source below cutoff | explains missing authoritative context | moving it inside context makes correct answer plausible | consumer dominance remains possible | SUPPORTED_CAUSE | | Consumer source wins scope match | directly explains wrong number and scope | suppressing mismatch removes supplied support for bad claim | ordering loss overlaps but is distinct | SUPPORTED_CAUSE | ### BLOCK 6: EVIDENCE-RANKED DIAGNOSIS Primary cause: scope-insensitive ordering selected consumer evidence over an enterprise source — OBSERVED. Competing/linked cause: the enterprise source fell below the generation cutoff — OBSERVED. Confidence state: DIAGNOSIS_SUPPORTED. ### BLOCK 7: MINIMAL REPAIR Current mechanism → fused candidates ordered without customer-scope precedence. Proposed change → add enterprise/consumer scope metadata as a deterministic final-ordering feature before cutoff. Failure class → WRONG_CHUNK_SELECTED / ORDERING_LOSS. Trade-off → requires reliable scope metadata. Expected effect → PROJECTED; the enterprise source should enter generation context for enterprise queries, not measured until replayed. Deferred → embedding replacement and re-chunking; neither targets the observed stage. ### BLOCK 8: REPLAY SPECIFICATION Replay the exact query. Capture complete candidates, scope metadata, pre/post ordering, generation cutoff, final context, answer, and citation. Confirm if the enterprise source moves inside the cutoff and the consumer rule is not applied. Falsify the repair if the enterprise source remains absent or the answer still copies the consumer rule despite correct context. ### BLOCK 9: ANALYSIS LIMITS The live pipeline was not replayed. This analysis does not establish general retrieval quality or production causality beyond the supplied incident.