---
name: Seed — Substrate, Governance, and the Limits of Constitutional AI
description: Major thinking document seed authorized 2026-03-22. Requires full corpus reading (David + Seb thinking folders) before writing. Three sections: relational mapping, L2 pressure audit, lines of inquiry toward a better approach. Not "better LLM" — possibly a different approach to non-human intelligence entirely.
type: project
---

# Seed: Substrate, Governance, and the Limits of Constitutional AI

**Filed:** 2026-03-22
**Status:** COMPLETE — written 2026-03-24, 33KB document
**Output path:** `~/_Dev/CapableMind-AI/docs/thinking/David/substrate-governance-inquiry.md`
**Tag:** [PROPOSAL] — thinking document, no code changes

## Central Question

> Are the goals of CapableMind achievable using LLMs as they are conceived of today? Or is the next step — once the system functions at a limited level — to build LLMs with this very problem in mind?

David's clarification: perhaps not a "better" LLM but a different approach to building non-human intelligence entirely. "Let your reverie take you to unexpected places."

## Required Corpus Reading (MUST complete before writing)

1. `~/CLAUDE.md` — global governance, executor agency directives, contamination awareness
2. ALL documents in `~/_Dev/CapableMind-AI/docs/thinking/David/` — every subdirectory
3. ALL documents in `~/_Dev/CapableMind-AI/docs/thinking/Seb/` — full thinking folder
4. `contamination-problem.md` in full — load-bearing for entire inquiry
5. L1 reliability artifacts in `docs/thinking/David/l1-reliability/`
6. Any L2 constitutional documents in the repository
7. Local CLAUDE.md files for build state and terminology

## Three Sections (related but distinct — do not collapse)

### Section 1: Relational Mapping
How does the central question land against current CapableMind work? What does the corpus already know? Where does L1/L2 architecture address the substrate question — where does it assume it away? Does the Harrison/Vico lens expose gaps in the contamination problem's framing?

### Section 2: L2 Pressure Audit
What is L2 designed to do and not do? Where could L2 push harder against contamination without training-level intervention? What mechanisms are within reach but not yet specified? Name what is genuinely out of reach. Produce concrete [PROPOSAL] candidates.

### Section 3: Lines of Inquiry — Toward a Better Approach
Speculative. Epistemic status marked explicitly. What would training with contamination in mind look like? Interpretability frontier? Human/AI collaboration if substrate problem substantially reduced? Precedents outside AI — institutional design, jurisprudence, contemplative practice, history of tool-making? Hold uncertainty honestly.

## Key Context from This Session

- The debugging itself was empirical evidence: executor accepted "non-fatal" at face value, didn't question 221-event assumption, proposed sequential fixes instead of stepping back. Each maps to contamination problem.
- Executor Agency Directives partially mitigated but are governance on top of substrate, not substrate change.
- Craig Mod: consciousness as "felt uncertainty," language as pointer never the thing itself
- LeCun/AMI Labs: JEPA, world models, $1B bet that language is wrong substrate
- Harrison/Vico on retention: retained framework shapes what can be seen, selective retrieval is function of what is already retained
- David's position: finish CapableMind at partial fidelity because naming limits precisely is precondition for designing what comes after
- "The Chamber doesn't need comprehension. It needs attention. Those might require fundamentally different architectures."

## Filing

- File completed document at output path
- Append [PROPOSAL] items from Section 2 to ~/PENDING.md
- Do not modify existing documents
- Do not implement architectural changes
- Report: document location, word count, number of [PROPOSAL] items, any escalation items