title: "Gather — "What do we NOT know?"" source: "tasks/TFW-25__values_consolidation/research/gather.md"
Gather — "What do we NOT know?"¶
Parent: HL-TFW-25 Goal: Validate TFW's 3-tier taxonomy and value count cap against external frameworks and industry patterns.
Findings¶
G1: Industry-standard Values/Principles/Rules hierarchy¶
Source: Software engineering taxonomy research (web search)
External consensus confirms a 3-layer hierarchy:
| Layer | Nature | TFW equivalent |
|---|---|---|
| Values | Abstract, foundational, "why" — culture, priorities | README §Values |
| Principles | Heuristic, guiding, "how" — best practices, design rules | KNOWLEDGE §0 (some) |
| Rules | Concrete, prescriptive, "what" — enforceable constraints | conventions.md |
Key insight: "The common failure mode occurs when teams implement Rules without fostering the Values that make those rules meaningful." This validates TFW's approach of having a separate Values section — but also means values must be genuinely philosophical, not disguised rules.
Implication for H1: P10-P13 (token density, inline enforcement, DNA/library, progressive disclosure) are clearly Rules-tier, not Values-tier. They describe enforceable constraints, not beliefs. The HL's proposal to move them to conventions.md is correct per industry taxonomy.
G2: Framework comparison — how many values/principles?¶
Source: Cursor rules, CrewAI, Microsoft RA, general AI framework docs
| Framework | Values-tier items | Format | Location |
|---|---|---|---|
| Cursor rules (2025 best practices) | 3-5 "always apply" rules | YAML frontmatter + markdown | .cursor/rules/ |
| CrewAI | ~4 core design principles (80/20, specialists, single purpose, flows vs crews) | Prose | docs site |
| Microsoft Responsible AI | 6 principles (fairness, privacy, transparency, accountability, reliability, inclusiveness) | Narrative sections | corporate docs |
| NIST AI RMF | 4 core functions | Structured prose | gov standard |
Pattern: Mature frameworks have 4-8 values-tier items. None exceed 10. All use narrative/prose, not tables. This strongly supports H3 (≤8 items).
Additional pattern — "Reference, Don't Duplicate": Both Cursor and general AI framework docs explicitly recommend referencing canonical files rather than duplicating content. TFW already follows this (P2), validating that it's a real value, not invented.
G3: Cursor rules — modular architecture principle¶
Source: Cursor rules best practices 2025
Key principle: "Keep individual rule files under 500 lines. Every token used for rules is a token taken away from the AI's ability to 'see' your actual code."
This validates TFW's P10 (token density) and P13 (progressive disclosure) — but as engineering constraints, not philosophical values. They are tools-implementation rules. In the Values/Principles/Rules taxonomy, they sit firmly in Rules.
G4: "Embed values into architecture, not just documents"¶
Source: AI agent framework best practices 2025-2026
Trend: By 2026, the shift is from "creating values docs" to "embedding values into agent architecture." Compliance = technical feature, not policy document.
This is exactly what TFW does with conventions.md, workflows, and templates. The README Values section should state the beliefs; the enforcement lives in the architecture. This supports the HL's approach: values = beliefs (README), rules = enforcement (conventions).
Checkpoint¶
| Found | Remaining |
|---|---|
| 3-tier taxonomy validated by industry standard | None — taxonomy is well-supported |
| 4-8 items is the norm for values sections | None |
| Narrative format preferred over tables for values | None |
| P10-P13 confirmed as Rules, not Values | None |
Sufficiency: - [x] External source used? (4 web searches, multiple framework references) - [x] Briefing gap closed? (taxonomy validated, count validated, format validated)
Stage complete: YES → User decision: ___