P3D Pack 1 Phase 5 — Dry-Run Prompt rev2 Field-Adaptive Patch Report

Date: 2026-05-11 Author: Opus 4.7

---

Status flags

phase5_dryrun_prompt_rev2_status=PATCHED
semantic_field_registry=true
deterministic_resolution=true
no_fuzzy_selectors=true
no_column_names_from_memory=true
no_composition_interpretation_by_agent=true
bounded_function_output=true
ambiguous_field_handling=true
requires_GPT_User_review_before_dispatch=true

Paths

Prompt rev2: ...prompts/p3d-pack1-phase5-readonly-dryrun-tac-to-iu-migration-prompt.md (rev 2)
Report:      ...reports/p3d-pack1-phase5-dryrun-prompt-rev2-field-adaptive-patch-report.md

Top 4 fixes from rev1

1. §0.1 Semantic candidate field registry — 11 concepts (address_field, content_body, content_hash, provenance_profile, parent_ref, sort_order, section_type, lifecycle, publication_link, title_field, description_field), each with candidate labels + listed tables. Deterministic resolution: 0=ABSENT, 1=USE, >1=AMBIGUOUS (reported, not silently picked). No "or similar."

2. G6 no longer assumes body/content_hash/content_profile — uses concept IDs content_body, content_hash, provenance_profile. If any concept is ABSENT on a table → that sub-query skipped. If AMBIGUOUS → reported.

3. G7 no composition interpretation — agent lists ALL live composition_level values from entity_species, groups species by composition_level, presents ALL groups as candidates. Agent does NOT say "simplest" or "allows containment." GPT/User interprets via Điều 0-B.

4. G8/G9 bounded output — function analysis limited to: identity_args, parameter names, return type, trigger inventory. G9 extracts NEW.<field> references from gate source, checks each against target schema, labels SATISFIED/GAP/UNKNOWN. No agent interpretation of function behavior.

What triggered this

rev1 still had:

• canonical_address or similar (fuzzy)
• body, content_hash, content_profile as assumed field names
• "simplest composition" agent interpretation
• "analyze source" open-ended function analysis

All eliminated in rev2 via semantic field registry + deterministic rules + bounded output.

Honest pattern acknowledgment

This is the pattern across Phase 4B-5: I fix hardcode at one level, leave it at the next level down. Species names → fixed. Column names → caught again. Semantic assumptions → caught again. The root cause: I write from knowledge instead of from discovery. The semantic field registry forces the prompt to always ask PG "what columns do you have?" before referencing ANY field. This should break the pattern at the structural level, not just the instance level.

Confirmation

phase5_dryrun_prompt=REV2_READY_FOR_GPT_REVIEW
next_action=GPT_REVIEW_PHASE5_DRYRUN_PROMPT_REV2