title: "Extract — "What do we NOT see?"" source: "tasks/TFW-38__quality_enforcement/research2/extract.md"

Extract — "What do we NOT see?"¶

Parent: HL-TFW-38 Goal: Concrete review stage design, mode system, checklist decomposition, diagram index format.

Findings¶

E1: Review Stage Specification¶

Proposed 4-stage review flow:

COMPREHEND → VERIFY → ASSESS → SYNTHESIZE
   (read)     (audit)  (judge)   (decide)

Stage 1: Comprehend¶

Mindset: "Understand before judging." - Read RF, TS, HL (current Step 1 — no change needed) - Output: mental model of what was done, why, and how

Stage 2: Verify¶

Mindset: "Trust nothing. Check evidence." - Open 2-3 files from RF §1 → do the changes exist? - If RF §4 claims "tests pass" → verify test file exists, re-run if possible - Cross-reference RF §3 (AC) checkmarks against TS DoD - For docs tasks: verify deliverable files exist, check structure matches spec - Output: verification evidence (line numbers, command output, file existence)

Stage 3: Assess¶

Mindset: "Apply domain standards." - Run domain-relevant checklist items (universal + mode-specific) - Check RF §6-8 completeness - Judge philosophy alignment, style, standards - Output: checklist with evidence, not just ✅/❌

Stage 4: Synthesize¶

Mindset: "Decide and capture." - Write verdict (APPROVE / REVISE / REJECT) with rationale - Triage executor observations → tech debt - Capture fact candidates from review process - Update traces (Task Board, TECH_DEBT.md) - Output: complete REVIEW artifact

E2: Review Modes (paralleling research modes)¶

Like research has focused / deep, review needs domain modes that configure the Assess stage checklist.

Proposed mode structure:

# In PROJECT_CONFIG.yaml
tfw:
  review:
    default_mode: standard  # standard | content | analytical
    modes:
      standard:
        description: "Code/implementation tasks"
        checklist_items:
          - dod_met
          - code_quality
          - test_coverage
          - philosophy_aligned
          - tech_debt
          - security
          - breaking_changes
          - style_and_standards
          - observations_collected
          - rf_completeness
        verify_depth: "spot-check 2-3 files, re-run 1 test"
      content:
        description: "Writing, documentation, design spec tasks"
        checklist_items:
          - dod_met
          - content_quality  # replaces code_quality
          - source_verification  # replaces test_coverage
          - philosophy_aligned
          - tech_debt
          - style_and_standards
          - observations_collected
          - rf_completeness
        verify_depth: "verify deliverable existence, check structure"
      analytical:
        description: "Research output, analytical specs, positioning"
        checklist_items:
          - dod_met
          - analytical_quality  # replaces code_quality
          - source_attribution  # replaces test_coverage
          - philosophy_aligned
          - tech_debt
          - style_and_standards
          - observations_collected
          - rf_completeness
        verify_depth: "verify deliverable existence, check source citations"

Mode selection: Reviewer picks mode at Step 0 based on task type. Same pattern as research mode selection.

E3: Checklist Decomposition Into Universal + Mode-Specific¶

#	Check	Universal	Standard (code)	Content (writing)	Analytical
1	DoD met?	✅	✅	✅	✅
2	{quality}	—	Code quality	Content quality	Analytical quality
3	{coverage}	—	Test coverage	Source verification	Source attribution
4	Philosophy aligned	✅	✅	✅	✅
5	Tech debt	✅	✅	✅	✅
6	{domain gate}	—	Security	NDA compliance	—
7	{domain gate}	—	Breaking changes	—	—
8	Style & standards	✅	✅	✅	✅
9	Observations collected	✅	✅	✅	✅
10	RF completeness (§6-8)	✅	✅	✅	✅

Universal (all modes): #1, #4, #5, #8, #9, #10 = 6 items Mode-specific: #2, #3, #6, #7 = 2-4 items depending on mode

Total per review: 8-10 items. Within the <10 constraint budget.

E4: Stage Files vs Single REVIEW Artifact¶

Option A: Separate stage files (like research/)

review/
  comprehend.md  ← notes from reading
  verify.md      ← evidence collected
  assess.md      ← checklist results
REVIEW__*.md     ← final synthesis

Option B: Stages as sections within REVIEW (no separate files)

REVIEW__*.md
  §1. Comprehend (summary of what was reviewed)
  §2. Verify (evidence table)
  §3. Assess (checklist)
  §4. Verdict
  §5. Tech Debt
  ...

Analysis: - Research stage files serve a purpose: they accumulate over OODA loops and get synthesized. Review doesn't loop — it's a single pass through each stage. - Creating 3 extra files per review adds overhead for a single-pass process. - The REVIEW template already has sections that map to stages naturally. - BUT: the staged mindset is what matters, not the file structure. The workflow can enforce stages by ordering Steps differently and adding mindset blocks.

Verdict: Option B (sections in REVIEW) is better. The workflow enforces the cognitive mode transitions through step ordering and mindset blocks. The REVIEW artifact captures everything in one place. No extra files.

E5: Diagram Index Format for KNOWLEDGE.md¶

Proposed format for KNOWLEDGE.md §2 (Key Artifacts):

### Diagrams Index

| # | Description | Location | Type | Created |
|---|-------------|----------|------|---------|
| 1 | Auth flow sequence diagram | RF Phase H §8 | mermaid sequence | 2026-04-14 |
| 2 | RLS policy evaluation tree | RES iter1 Findings Map | ASCII | 2026-04-14 |
| 3 | ETL pipeline data flow | RF Phase F §8 | mermaid flowchart | 2026-04-13 |

Collection rule in docs.md: After each REVIEW with ✅ APPROVE, check RF §8 and RES Findings Map. If non-trivial diagrams exist, add reference to KNOWLEDGE.md §2 Diagrams Index.

E6: Word Budget Impact of Staged Review¶

Current review.md: 816 words.

Estimated staged review.md: - Step 0: Mode selection (like research Step 2): ~30 words - Step 1: Comprehend (current Step 1, minor additions): ~10 words added - Step 2: Verify (new — mindset block + verification actions): ~80 words - Step 3: Assess (current Step 2, reformulated with mode reference): ~20 words added - Step 4: Synthesize (current Steps 3-7 compressed): ~0 (restructure, no net addition) - Step 5: checklist item #10 addition: ~10 words

Total estimated: 816 + ~150 = ~966 words. Well within 1200 budget.

Checkpoint¶

Found	Remaining
4-stage model: Comprehend → Verify → Assess → Synthesize	None
3 review modes: standard, content, analytical	Naming needs validation
Stages as REVIEW sections, not separate files	None
6 universal + 2-4 mode-specific checklist items	None
Diagram index format for KNOWLEDGE.md	None
Word budget: ~966/1200 after changes	None

Sufficiency: - [x] External source used? (ISO verification/validation model applied in Gather) - [x] Briefing gap closed? (Concrete stage design, mode system, checklist decomposition complete)

Stage complete: YES → User decision: ___