Gather — Iteration 2¶

Parent: HL-TFW-41 Goal: Embed Zwicky heuristics into existing research stages naturally.

Findings¶

G1: GMA Step → TFW Stage Natural Mapping¶

GMA has 6 steps (from swemorph.com canonical source). TFW research has 3 stages + synthesis. The mapping:

GMA Step                          TFW Stage        Natural fit?
─────────────────────────────     ────────────     ────────────
1. Problem Decomposition          GATHER           ✅ Perfect — Gather already asks "what
                                                   do we NOT know?" Decomposing into 
                                                   independent parameters IS identifying
                                                   what you don't know.

2. Enumerate Values               GATHER→EXTRACT   ⚠️ Split — Gather collects raw options,
                                                   Extract structures them into parameters.

3. Construct Morphological Field  EXTRACT          ✅ Perfect — Extract is "what do we NOT
                                                   see?" Building the matrix is exactly
                                                   seeing the full space you didn't see.

4. Cross-Consistency Assessment   CHALLENGE        ✅ Perfect — Challenge is "what do we NOT
                                                   expect?" CCA = finding unexpected
                                                   inconsistencies between pairs.

5. Iterative Refinement           OODA loop        ✅ Already built-in — the OODA loop
                                                   within each stage handles this.

6. Synthesis & Evaluation         RES synthesis     ✅ Already built-in — RES template
                                                   has Decisions and Recommendations.

Key insight: The mapping is almost 1:1. GMA steps don't need a separate "Zwicky Box section" — they ARE the stage heuristics. The problem in HD-19 was that all 6 steps were crammed into Extract as one "Section C" blob, violating the natural stage roles.

G2: HD-19 Failure — Wrong Stage Distribution¶

In HD-19 extract.md, the Zwicky Box section appeared as one monolithic block: - Parameters defined in-place (should have been in Gather) - Values listed in-place (should have been in Gather findings) - "Рекомендуется" annotated before any elimination (should have been in Challenge) - No CCA at all (should have been Challenge's core job) - Everything crammed into Extract as a single table

What should have happened: 1. Gather finds G1: "Working schedule has 3 storage options" → these are dimensions with alternatives, not "findings" in the narrow sense. 2. Extract builds the full cross-reference matrix from Gather dimensions. No recommendations yet. 3. Challenge performs CCA: "D1-Alt1 × D2-Alt2 — compatible? Yes/No/Why?" and eliminates pairs. 4. RES picks from surviving configurations.

G3: The "Heuristic" vs "Instruction" Distinction¶

The iter 1 approach (DR4/DR5) was instruction-based: "MUST do Zwicky Box." The user observed this produces simulation.

A heuristic approach would GUIDE the researcher's thinking at each stage without naming "Zwicky" or "morphological analysis." The researcher naturally decomposes, enumerates, and eliminates because the template QUESTIONS lead there.

The mechanism: Each stage template's heading question already implies the right GMA step: - "What do we NOT know?" → You don't know the full parameter space → Decompose - "What do we NOT see?" → You don't see all combinations → Enumerate & cross-reference - "What do we NOT expect?" → You don't expect which combinations are inconsistent → CCA

The templates just need to make these implications EXPLICIT with one-line heuristics, not with procedural instructions.

G4: Existing Template Gaps — What's Missing¶

Current templates are maximally open-ended: - Gather: {content} — no guidance on HOW to structure findings - Extract: {content — patterns, comparisons, internal file analysis} — hints at patterns but no requirement to enumerate alternatives - Challenge: {content — what could go wrong, alternatives, counter-evidence} — mentions "alternatives" but doesn't require systematic elimination

What's needed: One structural heuristic per stage that nudges toward the GMA step without naming the methodology. The heuristic should feel like natural stage discipline, not like a separate exercise.

Checkpoint¶

Found	Remaining
GMA maps 1:1 to Gather→Extract→Challenge	Need to design exact template wording
HD-19 failure = wrong stage distribution (all in Extract)	Need to test proposed wording against HD-19
Heuristic > Instruction approach for preventing simulation	Need to compare with iter 1 DR4 approach

Sufficiency: - [x] External source used? (GMA canonical methodology from swemorph.com) - [x] Briefing gap closed? (Mapping complete, HD-19 diagnosed)

Stage complete: YES → Proceeding to Extract.