Skip to content

title: "Gather — Iteration 2: "What do we NOT know?"" source: "tasks/TFW-32__methodology_and_positioning/research2/gather.md"


Gather — Iteration 2: "What do we NOT know?"

Parent: HL-TFW-32 Predecessor: RES1 Goal: Address naming precision, visualization sections, business processes, multi-iteration enforcement.

Findings

G1: Naming precision — "Strategic Insight" vs "Knowledge Candidate"

External research confirms: Category names act as cognitive frames for LLMs. Specific names trigger specific model behavior, generic names trigger generic behavior.

Category Label LLM Behavior Pattern
"Knowledge" Lists facts. Focuses on breadth and coverage. Retrieval mode
"Strategic Insight" Prioritizes trends, causality, risks, recommendations. Synthesis mode
"Fact Candidate" Captures observations. Lists what was noticed. Reporting mode

Key finding: "Knowledge" = declarative/information-oriented (WHAT is the information). "Strategic Insight" = analytical/synthesis-oriented (WHAT DOES the information MEAN for our goals). These trigger fundamentally different cognitive modes in LLMs.

Critical problem with RES1 D2 ("Doc Candidates + Knowledge Candidates"): - "Knowledge Candidate" is too broad — agent behavior = dump everything that could be knowledge (generic coverage mode) - "Doc Candidate" is similarly vague — what kind of documentation? - RES1 chose names based on WORKFLOW ROUTING (which workflow processes this) rather than CAPTURE BEHAVIOR (what kind of thing to look for)

D28 analysis (Naming Creates Behavior):

Name What it tells the agent to LOOK FOR Precision
"Fact Candidates" (current) Observations, patterns, anything noteworthy Medium — too broad
"Doc Candidates" (RES1) Things about documentation? about the project? Low — unclear target
"Knowledge Candidates" (RES1) Anything that could be knowledge Very Low — maximally vague
"Strategic Insights" (current HL §11) Business context, stakeholder priorities, domain knowledge, decisions, motives High — specific target
"Observations" (RF §5, already exists) Technical patterns, tech debt, code quality issues High — specific target

I notice: TFW already HAS the precise names. They're just in different sections: - RF §5 Observations = operational/technical findings (agent-observed, internal) - HL §11 Strategic Session Insights = domain/stakeholder knowledge (human-sourced, external) - RF §6 / RES §FC Fact Candidates = catch-all for everything else

The problem isn't naming §11 — it's that §6 "Fact Candidates" is the vague one. "Strategic Insights" is already precise.

G2: Visualization/Diagrams — where they live in other frameworks

arc42 model (12-section architecture documentation):

Diagram Type arc42 Section TFW Equivalent
System Context §3 Context & Scope HL §3.1 (vision-level)
Container/Component §5 Building Block View RF (detailed result)
Runtime/Dynamic §6 Runtime View RF (sequence diagrams, API flows)
Deployment §7 Deployment View RF (infrastructure result)

Pattern observed: arc42 has diagrams in EVERY section, not one central place. But each diagram has a CLEAR PURPOSE at that level: - Vision/Context diagrams → planning artifacts (HL equivalent) - Detail/Runtime diagrams → result artifacts (RF equivalent) - Architecture Decision diagrams → ADR section (KNOWLEDGE equivalent)

C4 model hierarchy: Context → Containers → Components → Code. Each level has its own diagram. Maps to TFW: HL = Context level, RF = Container/Component level.

Key insight from external research: "Pragmatism — only document what is necessary. Only add diagrams when they provide more value than text."

Current TFW state: - HL §3.1 "Result Visualization" — EXISTS but is too narrow. Instructions say "show the outcome" (only one purpose) - RF — NO visualization section. All diagrams are ad-hoc, embedded wherever - RES — NO visualization section - TS — NO visualization section (but TS shouldn't need one — it references HL §3.1)

G3: Business processes — how to represent value delivery flows

BPMN + Value Stream Mapping in agile: - Two levels confirmed by external research: - High-level: Value Stream Map = flow from request → delivery of value (who, what, when). Used in PLANNING - Detail-level: BPMN process diagram = exact steps, decisions, API calls, DB operations. Used in RESULT documentation - "Keep documentation Just Enough and Iterative" — start with high-level in planning, add detail in execution - "Focus on handoff points (seams)" — where work moves between teams/systems - "Living artifacts" — diagrams update as process evolves

User input analysis (from briefing + HL S15): 1. "Бизнес процесс — как некий флоу от юзера и до получения ценности или результата" 2. "ИИ рисует аски арт или диаграмму последовательностей — там видно процесс, и понимает ли ИИ где его конец, где достигается ценность" 3. "Нам надо завести раздел со схемами — для дизайна/вижна, для бизнес процессов, для диаграмм" 4. "Везде должно быть одна общая глава с одним четким названием, которая на стадии знаний будет куда-то собираться"

I notice a tension: User wants ONE chapter name across all artifacts, but the PURPOSE of diagrams differs by artifact: - HL: "what it SHOULD look like" (vision) - RF: "what it ACTUALLY looks like" (result) - RES: "what we DISCOVERED" (findings visualization)

The content differs. But the CONSOLIDATION target is the same: during 📚 KNW, all diagrams from all artifacts should be collectable into one place in knowledge.

G4: Multi-iteration research — structural enforcement design

External research confirms: - "Checkpoint-Resume Pattern" = define critical moments, save state, resume from last known good state - "CLAUDE.md Pattern" = structured config/log file read at session start - The key is externalized workspace memory — files that both agents and humans can inspect/edit - "Aggressive Summarization" — store summaries, not raw history

VLM-3 actual experience (4 iterations): - User ran 4 separate research sessions - Each produced findings that changed assumptions from previous sessions
- No control file existed — user mentally tracked what was covered - This worked because user was actively involved, but would FAIL for delegated teams

THIS session (iteration 2) — what actually happened: 1. User read RES1, identified gaps and new insights 2. User came to new agent with: list of gaps, reference to RES1, explicit "create research2" 3. I (researcher) had to re-read: HL, RES1, all 4 stage files from research/, conventions, templates 4. Total context loading: ~8 files, ~60K characters 5. What was MISSING that would have helped: a structured summary of "what was covered, what was NOT covered, what changed during RES1"

What the researcher SHOULD output at the end (structural enforcement):

## Iteration Status
- Iteration: 1 of {min}
- Hypotheses tested: H1, H2, H3, H5, H8
- Hypotheses remaining: H6 (deferred), H7 (designed not tested)
- New hypotheses generated: none
- Gaps discovered: [list]
- Recommendation: Run iteration 2 focusing on [gaps + new hypotheses]
→ Launch next `/tfw-research` in a separate agent. Pass this file as context.

Coordinator gate in plan.md: - After RES: coordinator checks iterations.yaml - If current_iteration < min_iterations → MUST launch next research - If current_iteration >= min_iterations → coordinator CAN proceed to TS, OR launch more - Coordinator CANNOT skip to TS while iterations remain

Checkpoint

Found Remaining
"Strategic Insight" is ALREADY the precise name — D28 validates it Design: what to do with §6 "Fact Candidates" vagueness
arc42/C4 confirm: diagrams at EVERY level, not one place Design: exact section name, format, consolidation path
BPM confirms two levels: vision (HL) + detail (RF) Design: section template for each artifact
Multi-iteration: structural enforcement needed, checkpoint-resume pattern Design: exact format of iterations.yaml + researcher exit message
"Knowledge Candidate" triggers generic retrieval mode in LLMs → Keep "Strategic Insight", don't rename to "Knowledge"

Sufficiency: - [x] External source used? (AI naming research, arc42, C4, BPMN, VSM, checkpoint-resume patterns) - [x] Briefing gap closed? (all 4 hypotheses investigated)

Deep mode criteria: - [x] Hypothesis tested? H5b (naming: "Strategic" > "Knowledge" — confirmed), H6 (visualization: arc42 model gives concrete placement), H6b (business processes: two-level pattern), H7b revised (structural enforcement with checkpoint-resume) - [x] Counter-evidence sought? For H5b: sought reasons why "Knowledge Candidate" could work (it doesn't — triggers wrong LLM behavior). For H6: sought argument that ONE central section is enough (arc42 says no — each level needs its own).

Metacognitive check: I discovered something NEW — the existing TFW naming is ALREADY right. §5 Observations = operational (precise), §11 Strategic Insights = domain (precise), §6 Fact Candidates = the vague one that needs fixing. RES1 went wrong by trying to rename the precise names (Strategic → Knowledge) instead of fixing the vague one (Fact Candidates).

Stage complete: YES → User decision: ___