Skip to content

Session Intelligence Trust Model

Status: Draft Consumers: ao context assemble, ao context explain, ao codex start, future runtime relevance scoring

This contract defines which artifact classes are allowed to influence runtime behavior for Session Intelligence, and which classes stay discovery-only, experimental, or archive-only.

The goal is simple: do not let high-volume .agents capture turn into low-trust startup sludge.

Trust Tiers

Class Tier Default startup Planning Pre-mortem Post-mortem Notes
discovery-notes discovery-only no no no yes Brainstorms, raw discovery notes, and rough research are useful provenance, not default runtime context.
pending-knowledge discovery-only no no no yes Pending extraction artifacts are not stable enough for automatic injection.
raw-transcripts archive-only no no no yes Raw transcripts remain lookup-only because they are noisy and too large.
learning runtime-eligible yes yes no yes Ranked learnings may enter runtime payloads once they pass quality gates.
pattern runtime-eligible yes yes yes yes Patterns are compact enough to reuse when query-matched.
finding canonical yes yes yes yes Promoted findings are the highest-trust reusable runtime signal.
belief-book canonical yes yes yes yes Stable cross-domain doctrine promoted from healthy evidence.
playbook canonical yes yes yes yes Generated operator workflows from healthy topics and promoted packets.
knowledge-briefing runtime-eligible yes no no yes Preferred explicit goal-time briefing surface for a concrete goal; task-scoped, not universal policy.
planning-rule canonical yes yes yes yes Compiled planning rules are canonical prevention memory.
known-risk canonical yes yes yes yes Compiled pre-mortem checks are canonical risk memory.
next-work runtime-eligible yes yes yes yes Ranked next-work gives the next session continuity.
recent-session runtime-eligible yes no no yes Session summaries help startup recovery when query-matched.
research runtime-eligible yes yes no yes Research can help, but it does not outrank findings or compiled rules.
topic-packets experimental no no no yes Remain behind packet-health review before default runtime injection.
source-manifests experimental no no no yes Good provenance surfaces, poor default startup payloads.
promoted-packets experimental no no no yes Higher promise, but still behind health gates until rollout metrics stabilize.

Default Suppression Rules

The following classes remain suppressed from default startup context:

  • discovery-only artifacts
  • archive-only artifacts
  • experimental packet families

Suppression does not mean deletion. These classes remain valid lookup, provenance, and validation surfaces. The contract only restricts what enters automatic runtime payloads.

Runtime Ranking Implications

Session Intelligence should prefer this order when building startup or planning context:

  1. canonical artifacts
  2. matched knowledge briefings when the operator or runtime explicitly assembles goal-time context
  3. runtime-eligible artifacts
  4. experimental families only after explicit health review
  5. discovery-only or archive-only artifacts only by explicit lookup or post-mortem use

Explainability Requirement

Any runtime surface that injects context must be able to explain:

  • which artifacts were selected
  • why they were selected
  • which classes were suppressed
  • why packet families were considered missing, thin, or experimental

ao context explain is the first CLI surface that exposes this contract directly.