trace-first-starter

RES — TFW-25: Values & Principles Consolidation

Date: 2026-04-04 Author: Researcher (AI) Status: 🔬 RES — Complete Parent HL: HL-TFW-25 Mode: Pipeline


Research Context

TFW accumulated 14 principles (P1-P14), 33 architecture decisions (D1-D33), and 29 knowledge facts across 24 tasks. These live in 3 places (README Values, KNOWLEDGE §0, knowledge/ topic files) with heavy duplication and no clear hierarchy. This research validates the proposed consolidation approach: what taxonomy to use, what to keep, what to prune, and how to restructure README Values as the philosophical face of TFW.

Briefing

See research/briefing.md. Three hypotheses tested. Focused mode: 1 OODA loop per stage. User direction: philosophy facts = keep, narrative format stays, no missing principles. “Naming Creates Behavior” confirmed as a philosophical belief (linguistic relativity analogy).

Decisions

# Decision Rationale
R1 3-tier taxonomy (Values / Design Principles / Rules) is correct Validated by industry standard (Values = why, Principles = how, Rules = what). Middle tier (KNOWLEDGE §0) earns its existence: 7 items too specific for README, too strategic for conventions
R2 README Values: 8 items (was 5). Add “Traces Over Code”, “Structural Enforcement”, “Naming Creates Behavior”. Rename “Determinism and Safety” → “Honesty Over Convincingness” External frameworks: 4-8 values is the norm. “Traces Over Code” was in §Thesis but missing from §Values. New values sourced from P14, D28. Rename: current text is rules, not beliefs — rewrite as genuine value statement
R3 KNOWLEDGE §0: 7 items (was 14). Remove P4, P6, P10-P13. Keep P1-P3, P5, P7-P9 (compressed). Promote P14 to README P4/P6 = obvious from code. P10-P13 = engineering rules (token limits, inline patterns, DNA/library, progressive disclosure) — belong in conventions, not philosophy. P14 = philosophical belief about structural enforcement
R4 knowledge/ facts: 18 facts (was 29). Prune 11 self-evident facts (6 convention, 5 process). Keep all philosophy + constraint Self-evidence test: “Is this now obvious from the code it describes?” Convention F4/F6/F8-F10/F12 and Process F2/F3/F8-F10 are all demonstrated by the artifacts they describe. Exception: convention/F5 (ref-inside-step) kept — the named pattern has value beyond code (D28)
R5 KNOWLEDGE §3 Legacy: prune 18 of 35 items (all pre-TFW-22, fully resolved) Historical items from TFW-2 through TFW-15 era. Replacements fully implemented. No active references. Keep TFW-22 through TFW-24 items (recent, may still be referenced)
R6 KNOWLEDGE §4 Tech Stack: remove entirely 4 lines of trivially obvious content (Markdown, Git, IDE tools, MIT license). Derivable from repo files
R7 “Determinism and Safety” → “Honesty Over Convincingness” — rename, not remove External research: safety belongs at both value and rule layers. Current README text is 5 implementation rules, not a belief. Rewrite as value: “AI that sounds confident while being wrong > AI that refuses to answer.” Rules stay in conventions §12

Open Questions

# Question Status Answer
Q1 Where do P10-P13 land after removal from §0? Open HL proposes conventions.md §X “Design Rules” or integrate into §11. TS should decide exact placement

Hypotheses (from HL §10)

# Hypothesis HL Status RES Status Evidence
H1 P10-P13 are implementation rules, not principles — they belong in conventions.md open ✅ confirmed Industry taxonomy: all 4 are “Rules” tier (concrete, prescriptive, enforceable). Token limits, inline patterns = engineering constraints
H2 ≥5 knowledge/ facts are now self-evident from code and can be safely removed open ✅ confirmed (11) 11 facts pass self-evidence test. Code/templates/workflows demonstrate them. Info loss risk verified per fact in Challenge
H3 README Values section should stay under 8 items to maintain narrative impact open ✅ confirmed (exactly 8) External benchmarks: Microsoft RA = 6, CrewAI = 4, Cursor = 3-5. No mature framework exceeds 10. 8 = upper bound of healthy range

HL Update Recommendations

# What to update Source
U1 §3 Target: update value count from 6-8 → exactly 8. Add “Honesty Over Convincingness” (renamed, not removed) Challenge C1
U2 §3 Target: knowledge/ fact count from ~20 → 18 (adjusted from 12 prunable to 11) Challenge C3
U3 §4 Phase A step 1: include “Honesty Over Convincingness” rewrite alongside other README changes Challenge C1
U4 §5 DoD item 1: update range from “6-8 values” → “8 values” R2
U5 §5 DoD item 6: update from “≤22 facts” → “18 facts” R4
U6 Phase B: add step for P10-P13 placement in conventions.md — needs exact section decision Q1

Fact Candidates

Reviewing conversation history — the user’s key input was the philosophical grounding for “Naming Creates Behavior” (linguistic relativity / Sapir-Whorf analogy). This is already captured in process/F5. No new strategic facts emerged from this research beyond what’s already in the knowledge base.

# Category Candidate Source Confidence
FC1 philosophy “Honesty Over Convincingness” framing: AI confidence without correctness is more dangerous than refusal. This is TFW’s safety belief — distinct from the implementation rules in conventions §12 Challenge C1 + external research High
FC2 convention Mature AI frameworks universally use 4-8 items for top-level values/principles. None exceed 10. Narrative format (heading + paragraph) preferred over tables for values sections Gather G2 (Microsoft RA, CrewAI, Cursor, NIST) High

Conclusion

This research validated the HL’s core proposal (3-tier taxonomy, pruning strategy) with two significant corrections: (1) “Determinism and Safety” should be renamed and rewritten as a genuine value (“Honesty Over Convincingness”), not removed — external research confirmed safety belongs at the values layer; (2) convention/F5 (ref-inside-step pattern) should be kept because the named pattern has value beyond its code implementation (D28 applies). The industry-standard Values/Principles/Rules hierarchy maps precisely to TFW’s README/KNOWLEDGE/conventions structure, giving confidence that the consolidation will produce a well-organized, maintainable framework. Final numbers: 8 README values, 7 KNOWLEDGE §0 principles, 18 knowledge/ facts, 17 Legacy items.


*RES — TFW-25: Values & Principles Consolidation 2026-04-04*