← Back to docs

// Trust model

Trust model

Code and tests stay above summaries, wiki pages, handoffs, and external memory. Trust labels tell agents how much rechecking is required.

// source of truth

Source-of-truth order

.knowledge is a trust and routing layer, not a replacement for reading current code.

Agents should treat current code and current tests as stronger than generated summaries, wiki pages, session notes, or external memory.

  • Current code
  • Current tests
  • .knowledge/evidence/*.json
  • .knowledge/modules/*.json
  • .knowledge/decisions.json
  • .knowledge/wiki/*.md
  • .knowledge/sessions/*
  • External memory, advisory only
Quick Start PROOF LINK

Quick Start

The installed entrypoint teaches agents the trust order.

Trust report tooling PROOF LINK

Trust report tooling

Trust status is generated from repository-local artifacts.

// labels

Trust labels

Trusted and near-trusted knowledge can route the agent, but critical behavior still requires targeted source reads.

Suspect, stale, low-confidence, or needs-recheck knowledge is context only until source evidence confirms it.

  • trusted: usable for routing and limited planning
  • near_trusted: usable after targeted checks
  • routing_trusted: route only
  • advisory_only: context only
  • suspect, needs_recheck, low_confidence: re-read source before relying

// technical notes

Related research notes