P3D Pack 1 Phase 4C — Prompt rev3 Field-Adaptive No-Fuzzy Patch Report

Date: 2026-05-11 Author: Opus 4.7

---

Status flags

phase4c_prompt_rev3_status=PATCHED
gate0_scope_preflight_added=true
field_adaptive_select_star=true
no_assumed_column_names=true
no_fuzzy_selectors=true
deterministic_plausibility_labels=true
subordinate_exact_match=true
scale_brittleness_self_audit_updated=true
requires_GPT_User_review_before_dispatch=true

Top 4 fixes from rev2

1. GATE-0 scope preflight — 6 checks (table existence, registry rows, column existence, vocab key) that must ALL pass before any goal runs. If P5 (discriminator column) fails → discriminator goals BLOCKED, other goals continue.

2. G3 field-adaptive — SELECT * for data dump. Semantic field presence map derived by agent from actual columns + data types. No hardcoded field list (display_name, prefix, depth removed from output shape). Agent reports what columns EXIST, not what we expect.

3. G7 exact strategy match — "subordinate-like" language removed. Agent queries exact birth_code_strategy value of target_collection_subordinate, then finds collections with same exact value. Zero semantic interpretation.

4. G8/G9 deterministic labels — PLAUSIBLE / CONFLICT / UNKNOWN / N/A with explicit criteria table. PLAUSIBLE = management_mode matches governance_role (both live values, both exist). CONFLICT = both exist but don't match. UNKNOWN = field absent for comparison. No agent subjectivity.

What's preserved from rev2

• §0 SCOPE BLOCK (4 identifiers + 1 vocab key)
• §2 AUTHORITATIVE SOURCES table
• G1-G10 goal structure and intent
• §5 CONSTRAINTS
• §6 SCALE BRITTLENESS self-audit (expanded to 9 checks)
• Read-only/dry-run/no-write boundaries
• candidate_not_approved labelling

Confirmation

phase4c_prompt=REV3_READY_FOR_GPT_REVIEW
next_action=GPT_REVIEW_PHASE4C_PROMPT_REV3