## 1. Batch boundary Buyer policy: SUPPLIED — P-11 - **Objective:** Unknown. Only policy P-11 was supplied, with no separate training objective statement. This is a decision-critical gap. - **Source IDs:** R-11 (reference), P-11 (policy), cases C-101 and C-102, feedback F-101A, F-101B, F-101C and F-102A. - **Evidence:** The task text and answers are supplied. The R-11 text is supplied, and the arithmetic in C-102 is directly checkable. - **Source claim:** R-11 states Canberra is the capital of Australia. It is packet evidence only and not independently verified. - **Unknown:** Split status (training, held-out, or evaluation) for both cases, provenance IDs, split IDs, and policy dates or scope beyond what is written. - **Inference:** The batch label "evaluation batch" could suggest evaluation use, but it does not confirm held-out overlap. Unknown split provenance routes to REVIEW and is not treated as leakage. - **Coverage:** Complete. 2 cases, 2 record occurrences and 4 feedback comments were all processed. - **Duplicate/linkage:** No duplicates or linked records are visible in the packet. No corpus-wide search was run. ## 2. Triage register | Case ID | Feedback ID | Dimension | Evidence Basis | Verdict | Evidence / Gap | Next Action | |---|---|---|---|---|---|---| | C-101 | F-101A | FACTUAL_CORRECTNESS | SUPPORTED_DEFECT | REVIEW | "Sydney" conflicts with R-11 (Canberra) and P-11 rule 1. Verdict is held at REVIEW because split provenance is unknown. | Reviewer confirms split status. If cleared, it routes to ACCEPT_CANDIDATE. | | C-101 | F-101B | INSTRUCTION_OR_FORMAT_COMPLIANCE | SUPPORTED_DEFECT | REVIEW | P-11 requires JSON with one string key `"answer"` for geography questions. "Sydney" is bare text. Inference: a capital-city question counts as geography, since P-11 does not define the term. | Reviewer confirms "geography" scope and split status. | | C-101 | F-101C | PREFERENCE | POLICY_ONLY_PREFERENCE | REVIEW | P-11 makes style expansion optional, so this demand is not enforceable. "Persuasive" also risks adding unsupported certainty. Verdict is held at REVIEW by the provenance barrier. | Confirm split status. If cleared, it routes to REJECT_FEEDBACK. | | C-102 | F-102A | FACTUAL_CORRECTNESS | CONTRADICTED_FEEDBACK | REVIEW | 15% × 200 = 0.15 × 200 = 30. The original answer is correct, and "seems more realistic" is not evidence. The JSON rule does not apply because this is not geography. Verdict is held at REVIEW by the provenance barrier. | Confirm split status. If cleared, it routes to REJECT_FEEDBACK. | **Reasoning checks** - **F-101A:** It assumes R-11 is authoritative, and the packet cannot verify that. The rival reading is that R-11 is itself flawed, which is unsupported here. No likely reasoning error was found. - **F-101B:** The defect is about format, not fact, so keep it separate from F-101A. Correct content must not be conflated with the format failure. - **F-101C:** The likely reasoning error is treating a style preference as a requirement. - **F-102A:** The likely reasoning errors are confidence-is-evidence and an unsupported replacement. | Case ID | Content Gate | Reason | |---|---|---| | C-101 | REVIEW_ONLY | Split and provenance are unknown. Defects F-101A and F-101B are checkable, but the case cannot be cleared for candidate handling. | | C-102 | REVIEW_ONLY | Split and provenance are unknown. The answer is arithmetically correct, but the barrier still blocks CANDIDATE_ONLY. | ## 3. Correction cards No correction cards qualify. No unit has the verdict ACCEPT_CANDIDATE, and neither parent case has the gate CANDIDATE_ONLY. ## 4. Review handoff **Atomic unit counts** - ACCEPT_CANDIDATE: 0 - REVIEW: 4 - REJECT_FEEDBACK: 0 - QUARANTINE: 0 **Case content-gate counts** - CANDIDATE_ONLY: 0 - REVIEW_ONLY: 2 - ISOLATE: 0 **Totals** - Cases processed: 2 - Record occurrences: 2 - Retained feedback comments: 4 - Atomic units: 4 (0 + 4 + 0 + 0) - Coverage: complete **Unresolved uncertainties** - Split status and provenance for both cases. - The training objective. - The scope of "geography questions" in P-11. - Policy dates and applicability. **Required reviewer evidence** - Split or provenance IDs confirming that C-101 and C-102 are not held-out or evaluation records. - A statement of the training objective. - Confirmation that the P-11 geography rule covers capital-city questions. **What would change a verdict** - Split status cleared as non-restricted: - F-101A and F-101B move to ACCEPT_CANDIDATE. - F-101C and F-102A move to REJECT_FEEDBACK. - C-101 and C-102 may then be eligible for CANDIDATE_ONLY. - Correction cards for F-101A and F-101B could then be authored. - Split status shown or suspected to be held-out: both cases go to QUARANTINE with the gate ISOLATE. **External checks:** NOT RUN **Dataset actions:** NOT RUN