Research: context-sensitive axiom amplification — epigenetic tagging for task-type routing

[AccessiT3ch]

Research Question

How should the endogenic system amplify different axioms for different task contexts — and what is the minimal viable mechanism to implement this without new infrastructure?

Context

The current system treats all three axioms as equally active in all contexts. The values-encoding synthesis (H5 / Gap identified) notes this as a gap: no equivalent of DNA regulatory regions (promoters/enhancers) exists in the endogenic substrate.

The biological analog: in different cell types, different genes are amplified by regulatory regions even though all cells carry the same genome. The endogenic equivalent:

• Research/scouting tasks → amplify Endogenous-First (fetch before act; read prior scratchpads before scouting)
• Commit/release tasks → amplify Documentation-First (Documentation-First principle)
• Script-writing tasks → amplify Algorithms Before Tokens (programmatic-first; encode the solution)
• Local inference tasks → amplify Local Compute-First

This is sometimes called "epigenetic tagging" for the instruction substrate.

Scope

Survey three possible implementation mechanisms and produce a recommendation:

1. Agent-file metadata — a YAML frontmatter field amplify: [endogenous-first] that the VS Code mode/instruction layer uses to add emphasis to the relevant axiom block at session start
2. Task-type selectors in AGENTS.md — a lookup table mapping task keywords to axiom amplification hints, referenced in the session-start ritual
3. Context amplifier script — a scripts/amplify_context.py that prepends the relevant axiom block in amplified form to the session scratchpad start based on detected task type

Also survey: does Constitutional AI's per-task constitution variant (Bai et al. 2022) provide an applicable pattern here?

Gate Deliverables

• D1 — Comparison of three mechanisms (metadata / AGENTS.md selector / script); recommendation with rationale
• D2 — Proof-of-concept implementation of recommended mechanism in ≥1 agent file or AGENTS.md
• D3 — Findings documented in docs/research/values-encoding.md §5 (OQ-VE-2 resolution) or new section
• D4 — If a script is produced: tests with ≥80% coverage

Related Issues

• Source research: docs/research/values-encoding.md §5, OQ-VE-2; H5 gap ("regulatory regions")
• Companion: OQ-VE-1 (drift detection — detection without amplification is incomplete)
• Milestone: Value Encoding & Fidelity

[AccessiT3ch]

Phase 6 Complete — Epigenetic Tagging / Context-Sensitive Amplification

Phase 6 research and proof-of-concept for this issue is complete. Deliverables committed in PR #92 (research/skills-decision-logic → feat/value-encoding-fidelity), commit fd20524.

What was delivered:

• AGENTS.md — ### Context-Sensitive Amplification section added (the adopted POC mechanism)
• docs/research/values-encoding.md — OQ-VE-2 marked RESOLVED with mechanism comparison + rationale

Key findings:

• Three mechanisms evaluated: YAML frontmatter amplify: field, AGENTS.md lookup table, scripts/amplify_context.py
• Recommended & adopted (Phase 1): AGENTS.md task-type lookup table — single source of truth, zero new tooling, all agents already read AGENTS.md at session start
• Deferred (Phase 2): scripts/amplify_context.py — per Programmatic-First criterion, the table must be validated across ≥2 sessions before scripting the automation
• YAML frontmatter approach rejected: dispersed drift risk (15+ independent update points)

Gate deliverables status:

• D1 — Mechanism comparison + recommendation (AGENTS.md lookup table)
• D2 — POC implementation in AGENTS.md (### Context-Sensitive Amplification section)
• D3 — Findings in docs/research/values-encoding.md §5 OQ-VE-2 (RESOLVED)
• D4 — scripts/amplify_context.py — deferred to Phase 2 (after table validated across sessions)

Next: PR #92 awaiting review and merge. Phase 2 script (amplify_context.py) to be authored once the table is confirmed effective across ≥2 sessions.

[AccessiT3ch]

Phase B complete — research doc committed.

Artifact: docs/research/epigenetic-tagging.md (Status: Final)
Branch: research/bubble-clusters-phase-a-neuroanatomy — commit 20faefd

Coverage:

• Formal theoretical synthesis and mechanism comparison for OQ-VE-2 (values-encoding.md §5)
• Three mechanisms compared: YAML frontmatter / AGENTS.md lookup table / amplify_context.py script
• AGENTS.md lookup table adopted as Phase 1 mechanism (centralized, zero new tooling, single-edit propagation)
• Doc cross-references AGENTS.md §Context-Sensitive Amplification as the implementation artifact; research frames the theoretical basis (epigenetics, Constitutional AI per-task variants)
• Phase 2 script design (deferred pending two-session validation of Phase 1)
• 2 patterns (F1: Centralized Epigenetic Register, F2: Retrieval-Augmented Amplification Script) with canonical examples and anti-patterns
• 3 recommendations

Validation: validate_synthesis.py PASS, pytest 425 passed.