Skip to content

title: "Challenge — Iteration 4: "What do we NOT expect?"" source: "tasks/TFW-32__methodology_and_positioning/research4/challenge.md"


Challenge — Iteration 4: "What do we NOT expect?"

Parent: HL-TFW-32 Predecessors: Gather, Extract Goal: Stress-test per-template naming, visual concept split, and the "Value Flow" name.

C1: Does per-template naming confuse new agents?

Concern: If HL has "Value Flow" and RF has "Diagrams" — a fresh agent reading both for the first time might not realize they're the "same kind of section" with different framing.

Counter-test scenarios:

Scenario What happens with unified "Diagrams"? What happens with per-template names?
Agent writes HL Sees "Diagrams" → thinks of UML/ERD → writes tech diagrams in a vision document Sees "Value Flow" → thinks of user journey, value delivery → writes strategic overview ✅
Agent writes RF Sees "Diagrams" → draws tech diagrams ✅ Sees "Diagrams" → draws tech diagrams ✅
Agent writes RES Sees "Diagrams" → draws... what? Research findings as UML? Unclear Sees "Findings Map" → draws concept map, comparison table, decision tree ✅
Agent collects from all 3 Knows to look for "Diagrams" in each file Needs to know: HL = "Value Flow", RF = "Diagrams", RES = "Findings Map"

Key finding: For the WRITER agent, per-template names are BETTER. Each name prompts the right cognitive mode. For the COLLECTOR agent, it's slightly more complex — but the collector reads the whole file anyway.

The confusion risk is real only if agents need to cross-reference between templates by section name. In TFW, they don't — they read the template header, which tells them what to do.

Verdict: Per-template naming for visual section = net positive. Writer confusion DOWN (better prompts), collector cost = negligible.

C2: "Value Flow" as HL visual section name — stress test

Domain What does "Value Flow" prompt? Works?
Software product (e2e feature) User → input → processing → value received
Business analytics (CFO report) Data → analysis → insight → decision → business value
Education (curriculum design) Student → learning activity → skill acquisition → assessment → competence
Research project (TFW-32) Problem → investigation → finding → decision → methodology improvement
Operations (logistics) Order → fulfillment → delivery → customer satisfaction

Cross-domain: ✅ ALL work. "Value Flow" = universal. Every project delivers value through a flow.

Edge case: Pure research with no clear "value delivery" (e.g., "explore whether X is feasible"). → "Value Flow" still works: question → method → evidence → answer → decision. The "value" is: clearer decision.

Alternative considered: "Value Delivery" — too formal, sounds like a KPI metric, not a visual prompt.

Alternative considered: "Process Overview" — generic, doesn't prompt VALUE thinking, agent would produce boring boxes-and-arrows.

Verdict: "Value Flow" is strong. Cross-domain, value-centric (TFW-native), prompts the right level of abstraction for HL.

C3: "Findings Map" as RES visual section name — stress test

Scenario What does "Findings Map" prompt?
Research compared 5 tools Comparison matrix / radar chart ✅
Research investigated a hypothesis Decision tree: hypothesis → evidence → verdict ✅
Research gathered market data Market landscape map / positioning chart ✅
Research analyzed code architecture Component relationship diagram ✅

"Findings Map" = "organize and visualize what we found." The "Map" part prompts spatial organization (not just a list). The "Findings" part filters to research output (not process).

Counter-question: Is "Findings Map" too research-specific? RES is ONLY used in research. So yes, it should be research-specific. That's the point.

Verdict: Works well for RES context.

C4: HL §3.1 "Result Visualization" — does it still need to exist alongside "Value Flow"?

The question I didn't expect: If HL gets a "Value Flow" section, what's the difference between §3.1 "Result Visualization" and the new "Value Flow" section?

§3.1 Result Visualization NEW Value Flow section
Shows the OUTCOME — "what does done look like" Shows the PROCESS — "how does value get from A to B"
Amazon Working Backwards. Tables, examples, sample output Lean/BPMN tradition. Flowcharts, journey maps, architecture overview
Part of §3 "Target State (To-Be)" — embedded Separate section — standalone
ALWAYS present (required in HL) OPTIONAL (not every HL needs a process diagram)
TEXT-heavy: before/after tables, mockups, code snippets VISUAL-heavy: diagrams, flows, maps

These are genuinely different: - §3.1 = "Imagine the result exists. Show me." - Value Flow = "Show me how the machine works that produces that result."

Example from TFW-32 HL: - §3.1 shows: pipeline comparison (before → after), docs-vs-knowledge separation diagram, terminology unification table. These are OUTCOMES. - Value Flow would show: the workflow sequence (HL → RES → TS → RF → REVIEW → KNW → DONE), or the knowledge collection flow (Fact Candidates → /tfw-knowledge → verified facts). These are PROCESSES.

Both are valuable. Both are different. Both should exist.

Verdict: §3.1 "Result Visualization" and "Value Flow" are complementary, not overlapping. Keep both.

C5: The "generic section label" question for conventions.md

User asked: «в теории мы можем просто иметь один общий номер раздела и как-то его условно назвать»

What should conventions.md say about the visual section?

Option A: Named reference

### Visual Section (per-template)
Each result artifact has a visual section with template-specific naming:
- HL: "Value Flow" — value delivery overview
- RF: "Diagrams" — technical process detail
- RES: "Findings Map" — research findings visualization

Option B: Numbered reference

### §VIS (Visual Section)
Section number §VIS appears in HL, RF, and RES templates with template-specific names.

Option C: No special marking — just template instructions Each template defines its own section. No cross-cutting reference needed. Agents read their own template.

My assessment: Option A is cleanest. Conventions.md documents the pattern. Each template implements its own name. Collection algorithm reads by position, not by name.

But — do we even NEED a cross-cutting reference? The collector agent reads the full artifact. It doesn't search by section name. It reads everything.

The cross-cutting reference is useful for: 1. Human understanding: "where do diagrams go?" → "each template has a visual section" 2. Glossary entry: one place to see all visual section names 3. Template consistency: new templates know they should include a visual section

Verdict: Option A. Document the pattern in conventions.md. Brief, not heavy.

C6: Strategic Insights qualifier — final stress test

RES3 left Q8 open: should Strategic Insights have qualifiers?

With per-template naming established for visual section, let me re-check:

Template Without qualifier With qualifier
HL §11 "Strategic Insights" "Strategic Insights (Planning)"
RF §7 "Strategic Insights" "Strategic Insights (Execution)"
RES §7 "Strategic Insights" "Strategic Insights (Research)"

Does the qualifier change agent behavior? RES3 empirical test showed "Strategic Insights" triggers deep analytical mode. Adding "(Planning)" or "(Execution)" = context, not behavior change. The core prompt word is "Strategic Insights" — qualifier is parenthetical annotation.

Does the qualifier help the human? Yes. When reviewing: "is this from planning or execution?" matters. The qualifier answers without reading the whole artifact.

Does the qualifier help collection? No. Collection gathers ALL Strategic Insights. Source context is in the table (Source column).

Verdict: Qualifiers are useful for humans, neutral for agents. Add them. Cost = zero. Benefit = human clarity.

Checkpoint

# Finding Status
C1 Per-template naming for visual = net positive. Writer confusion ↓, collector cost ≈ 0 ✅ CONFIRMED
C2 "Value Flow" is cross-domain, value-centric, prompts right level ✅ CONFIRMED
C3 "Findings Map" works for RES context ✅ CONFIRMED
C4 §3.1 "Result Visualization" and "Value Flow" are DIFFERENT concepts. Both needed ✅ CONFIRMED: two sections, not one
C5 Conventions.md should document the visual section pattern (Option A) Recommend
C6 Strategic Insights qualifiers: useful for humans, neutral for agents. Add ✅ CONFIRMED

Sufficiency: - [x] External source used? (Gather G1-G5 informing all challenges) - [x] Briefing gap closed? (All 3 hypotheses challenged and resolved)

Deep mode criteria: - [x] Hypothesis tested? H_pertemplate (confirmed FOR visual, rejected for §6/§7), H_visionvs (confirmed: two separate concepts), H_valuemaps (rejected: "Value Maps" too loaded, "Value Flow" is better) - [x] Counter-evidence sought? Tested confusion risk (C1), cross-domain coverage (C2-C3), overlap risk (C4) - [x] Metacognitive check: The unexpected finding was C4 — I came in thinking "Value Flow" replaces §3.1, but actually they're complementary. §3.1 = WHAT done looks like. Value Flow = HOW the machine works. Both needed in HL.

Stage complete: YES


C7: Empirical LLM Validation (Post-Challenge)

Two experiments run against Qwen3.5-27B locally (vLLM, 192.168.1.109:8000).

Experiment 1: Same context (school management system), 6 different section names. temp=0.3, max_tokens=2000. Experiment 2: Same section name in different template contexts (HL vision vs RF result report). temp=0.3, max_tokens=1500.

Full analysis: experiment_analysis.md Raw data: captured in chat session (Qwen3.5-27B vLLM, 2026-04-10)

Experiment 1 Results: Each name triggers a DISTINCT cognitive mode

Section Name Cognitive Mode Content Type
Value Flow Strategic / Value-oriented Value streams, INPUT→PROCESSING→OUTCOME, "Value Created" column
Diagrams Technical / Engineering Mermaid (ERD, graph TD, swimlane), system architecture, schema
Process Maps Operational / BPM As-Is vs To-Be, pain point annotations, per-step timing
Findings Map Analytical / Research Root cause analysis, hypothesis tree, priority matrix
Visual Overview Generic / Unfocused Mixed architecture — no specific mode activated
Result Visualization Narrative / Outcome "6 Months After Launch", timeline (8:00 AM →...), testimonials

Key finding: Zero overlap between "Value Flow" and "Result Visualization". They are empirically distinct concepts.

Experiment 2 Results: Context × Name interaction

HL Context RF Context
"Value Flow" ✅ Clean strategic output, value principles ⚠️ OK but adds tech details
"Diagrams" ⚠️ CONFUSED — mixed vision/tech ✅ Pure technical diagrams

Key finding: "Diagrams" in HL = model confusion. "Value Flow" in HL = focused strategic output. Confirms per-template naming decision.

Hypothesis Updates

Hypothesis Analytical status Empirical status
H_pertemplate (visual section) CONFIRMED (C1) CONFIRMED — 6 modes + context test
H_visionvs (§3.1 ≠ Value Flow) CONFIRMED (C4) CONFIRMED — zero content overlap
H_valuemaps → Value Flow REFUTED (C2) CONFIRMED — "Value Flow" triggers right mode
Findings Map for RES CONFIRMED (C3) CONFIRMED — research-native output
Visual Overview = weak CONFIRMED — generic, unfocused
Diagrams wrong for HL CONFIRMED — model produces confused mix

→ ALL naming decisions now have BOTH analytical AND empirical validation. → Proceed to Synthesis (RES)