---
type: concept
topic: agent-wiki
status: active
created: 2026-04-07
updated: 2026-04-07
source_notes:
  - "[[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]]"
  - "[[research/agent-wiki/sources/karpathy-llm-wiki]]"
  - "[[research/agent-wiki/sources/anthropic-building-effective-ai-agents]]"
derived_from:
  - "[[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]]"
confidence: high
shared_concept:
---

# Governed Interpretation Layer

## Definition

A governed interpretation layer is a persistent wiki that accelerates synthesis and orientation, while explicitly refusing to become the unquestioned truth store. It stays useful by keeping canonical source notes read-only, preserving contradictions, and treating retrieval over authoritative notes as the verification backstop.

## Why It Matters Here

This is the trust boundary for the workspace. Without it, the compiled wiki risks becoming a plausible but increasingly self-referential cache of prior interpretations. With it, the workspace can compound insight without pretending that synthesis has replaced evidence.

## Supporting Sources

- [[research/agent-wiki/sources/dissenting-views-on-persistent-agent-wikis]] argues that persistent memory turns drift, poisoning, and provenance loss into durable system risks.
- [[research/agent-wiki/sources/karpathy-llm-wiki]] provides the positive case for a compiled markdown layer that persists beyond any one query.
- [[research/agent-wiki/sources/anthropic-building-effective-ai-agents]] supports the constraint that the surrounding workflow should remain inspectable and deliberately scoped.

## Tensions

- More governance improves trust, but it also adds friction to ingest and maintenance.
- A persistent wiki can be easier to reason from than raw retrieval, but easier reasoning is not the same thing as fresher truth.
- Retrieval corpora can also be poisoned, so governance is not an excuse to romanticize plain RAG.

## Implications

- Canonical source notes should remain read-only and serve as the final verification surface.
- Durable concept pages should keep visible links back to source digests rather than laundering evidence into standalone prose.
- Contradictions should be preserved and surfaced in `health-check.md`, not resolved cosmetically.
- Promotion should stay slow because promotion is a trust decision, not just a deduplication step.
- High-stakes answers should prefer a hybrid pattern: use the compiled layer for orientation, then retrieve against canonical notes for recency and verification.

## Related Concepts

- [[research/agent-wiki/concepts/compiled-knowledge-layer|Compiled Knowledge Layer]]
- [[research/agent-wiki/concepts/simplicity-before-autonomy|Simplicity Before Autonomy]]
- [[research/agent-wiki/concepts/obsidian-as-agent-ide|Obsidian as Agent IDE]]

## Related Outputs

- [[research/agent-wiki/outputs/wiki-vs-rag-comparison|Wiki vs RAG Comparison]]
- [[research/agent-wiki/outputs/first-pass-thesis|First-Pass Thesis]]