# Response to DeepReview — legal_contract_diligence_instance_1 (Revision 1)

Thank you for the careful review. Every finding is addressed below; where to verify each fix is stated inline.

**Change 1 — risk-selection determinism.** Resolved by removal rather than repair: the per-contract risk-summary deliverable is withdrawn from the task by author design decision. We state this plainly because the reviewer inspected the risk keys: the "highest-exposure" superlative was the one scored judgment the task asked for, and rather than constrain it with a priority rule and acceptance sets, the author elected to remove it, leaving the three deterministic layers (mapping, 90-cell extraction, portfolio questions). The retired risk keys and rationales remain archived; the self-test reports' risk sections stand as historical record. Verify: instructions v6.8 (no risk deliverable), bundle (13 files, no risk_summary.xlsx), spec §Deltas item 5.

**Change 2 — grader gaps.** The shipped score_task.py now compares amendment-mapping cells as sets (multi-value 'Amended By'), and its defaults match the shipped scale (5 contracts, 96% threshold). The risk-scoring gaps identified are mooted by Change 1's removal; the harness scores exactly the three layers the spec defines, and overall pass = L1 ∧ L2 ∧ L3. Smoke-tested end-to-end (perfect submission passes; order-flipped 'A02; A01' mapping passes; planted errors fail the correct layers). Verify: score_task.py (3-layer), scoring_config_N5.json, spec §Evaluation.

**Change 3 — template and provenance.** Field 7's validation list now includes `Redacted` (the frozen key's C02 value is enterable). The tracker Legend is rewritten: stale 36-contract scope, two internally contradictory C4 rows, contradictory C6 ordering rows, stale amendment-ID formats, and the C8 field-list omission are all removed; the sheet now summarizes the v6.8 rules only and states that diligence_instructions.docx controls on any conflict. Both provenance CSVs' A02 labels now read v2.3, matching the shipped amendment text. Verify: tracker_template.xlsx (Tracker H-column validation; Legend sheet), final_manifest.csv, amendment_filename_map.csv.

**Self-test evidence.** Headline results are unchanged by Revision 1 — 0/6 (Sonnet 4.6 reference band) and 0/4 (Opus 4.8 max effort) — because L2 extraction was the binding layer in every trial under both the old and new layer sets. The reports uploaded as Difficulty Self-Test evidence carry dated addenda noting the risk layer's removal; their risk sections are historical.