Skip to content

TASK_PHASE4_SALVAGE.md

Latest commit

 

History

History
79 lines (56 loc) · 2.58 KB

TASK_PHASE4_SALVAGE.md

File metadata and controls

79 lines (56 loc) · 2.58 KB
Raw

TASK: Phase 4 — Fragment Salvage + Refinement

Created: 2026-03-31 Status: IN PROGRESS Branch: development Project: MindSpark: ThoughtForge

Task Scope

Build the refinement layer: FragmentSalvage (multi-pass draft scoring + reassembly), EnforcementGate (citation integrity + response validation), wire both into ThoughtForgeCore, and deliver run_thoughtforge.py interactive CLI.

What Is Done (Phase 3 baseline)

  • Phase 0–3 complete, 191 tests passing
  • ThoughtForgeCore.think() full pipeline working
  • Inline fragment scoring in core.py (to be replaced/augmented)
  • Inline citation check in core.py (to be replaced with EnforcementGate)
  • refinement/init.py stub only

What Is Pending (Phase 4 Work)

  • Write this TASK file + commit
  • src/thoughtforge/refinement/salvage.py — FragmentSalvage.forge()
  • src/thoughtforge/refinement/enforcement.py — EnforcementGate
  • src/thoughtforge/refinement/__init__.py — exports
  • Update src/thoughtforge/cognition/core.py — wire salvage + enforcement
  • run_thoughtforge.py — interactive CLI
  • tests/test_phase4_refinement.py — test suite
  • Git commit Phase 4

Phase 4 Module Map

Module Class Key Method
refinement/salvage.py FragmentSalvage forge(candidates, bundle, engine, prompt_builder, sketch, scaffold) -> SalvageResult
refinement/enforcement.py EnforcementGate check(text, citations, bundle) -> EnforcementResult
run_thoughtforge.py — Interactive REPL + single-query mode

FragmentSalvage.forge() Logic

  1. Score each candidate: length_score (45%) + citation_score (55%) per BUILD_PLAN_v1
  2. Extract top sentences from best-scoring candidates
  3. Attempt refine pass via engine (if available) — max 2 passes
  4. Return SalvageResult: text, citations, confidence, passes_used

EnforcementGate.check() Logic

  1. If no knowledge retrieved → pass (nothing to enforce against)
  2. If Wikidata records present → check at least 1 QID cited
  3. If text below min length → fail
  4. Return EnforcementResult: passed, notes, text (may append forge note on soft fail)

Key File Locations

File Path
Build Plan Phase 4 BUILD_PLAN_v1.md lines 134–146
Scoring spec docs/specs/Retrieval_and_Scoring_Spec.md
ThoughtForgeCore src/thoughtforge/cognition/core.py
Refinement package src/thoughtforge/refinement/
CLI run_thoughtforge.py

Resumption Instructions

Read this file + check git log. Continue from first unchecked item above.