AI Agent Learning authority

Agent Learning is a project-isolated authority for evidence-backed procedural, failure-avoidance, relational, tool-routing, and host-compatibility lessons. It never becomes Project Truth.

Current counts are derived release facts, not permanent ceilings.

Action	Access	Internal owner
`learning_decide_candidate`	write-capable	`agent_learning:decide_candidate`
`learning_inspect`	read	`agent_learning:inspect`
`learning_record_host_memory_import`	write-capable	`agent_learning:record_host_memory_import`
`learning_retrieve`	read	`agent_learning:retrieve`
`learning_revoke`	write-capable	`agent_learning:revoke`
`learning_seal_candidate`	write-capable	`agent_learning:seal_candidate`

Two learning lifecycles

Learning surface	Admission	Reuse	Human decision	Project effect
Per-Delta Learning	`AUTO_ACCEPTED_DELTA_LEARNING` only at verified Delta exit	Bounded input to the next Delta entry	No individual Learning HIL	No Project pointer, Overlay, or accepted-ZIP effect
Full-PV consolidated weave	`PENDING_LEARNING_HIL` after the full-PV candidate is sealed	Accepted weave can inform later work	Separate six-way Learning HIL	Moves only the Learning pointer
Evidence-backed reusable lesson candidate	Explicit `learning_seal_candidate` route	Bounded retrieval after acceptance	Separate Learning HIL	Never becomes Project Truth

Automatic Delta Learning exists so verified procedural evidence from one row can inform the next row without forcing a human decision after every Delta. It remains distinct from the row’s auto-accepted sub-PV work receipt and from the later consolidated full-PV Learning weave.

Candidates remain unaccepted until the separate Learning HIL records the exact decision. Accepted Learning moves only the Learning pointer; revocation is append-only and does not erase historical evidence. Host MEMORY.md can be linked only through an explicit nonauthoritative provenance receipt.

Learning may inform later work through bounded retrieval. It cannot change a Project pointer, accept a Project proposal, alter Canon, replace Project Memory, or infer HIL from repetition or model confidence.

Source-bound workflow map

This page is projected from the same current executable snapshot as the rest of the documentation set. The map is deliberately two-directional: each horizontal district shows peer stages while vertical edges show ownership and state progression.

flowchart TB
    subgraph InputDistrict["Input and classification"]
      direction LR
      A["Evidence-backed lesson proposal"] --> B["Learning candidate classification"] --> C["Agent Learning authority"]
    end
    subgraph ExecutionDistrict["Selection and execution"]
      direction TB
      D["Separate Learning HIL"] --> E["Accept, reject, or research"] --> F["Evidence and scope validation"]
    end
    subgraph EvidenceDistrict["Evidence and outcome"]
      direction LR
      G["Learning decision receipt"] --> H["Bounded later retrieval or revocation"]
      G -. mismatch .-> I["Keep Project Truth unchanged"]
    end
    C --> D
    F --> G

Contract and readback

Phase	Current contract	Required readback
Input	Evidence-backed lesson proposal	Exact identity, provenance, and scope
Classification	Learning candidate classification	Owning schema, action, lane, skill, or authority
Owner	Agent Learning authority	One canonical implementation owner
Route	Separate Learning HIL	Condition-true ordered route with no hidden alias
Execution	Accept, reject, or research	Real execution or a visible fail-closed result
Validation	Evidence and scope validation	Hash, schema, authority-effect, and negative-case checks
Receipt	Learning decision receipt	Content-addressed result and provenance receipt
Downstream	Bounded later retrieval or revocation	Only the explicitly eligible next state
Failure	Keep Project Truth unchanged	No inferred HIL, candidate acceptance, or pointer movement

Canonical source owners

authorities/agent_learning/manifest.v1.json
skills/evi-learning/SKILL.md
schemas/actions/learning_seal_candidate.schema.json

Exact backend readback

Source contract	Bytes	SHA-256
`authorities/agent_learning/manifest.v1.json`	7742	`A23A8299F02569AEA1934BAA2B4D2A55BA018A157470757906F1C6C4D778F521`
`skills/evi-learning/SKILL.md`	10124	`42AB4B361F38BD7744078B647190D28FB642D1F701CC90D79C53BEA12913E8E5`
`schemas/actions/learning_seal_candidate.schema.json`	3782	`1701BFA110AD78A7AA299B333E0BBF0DA4E59A7E1478F7247BED1BE306F7C1E0`

Cross-surface invariants

The current snapshot contains 91 public actions, 26 skills, 11 hook events / 44 handlers, 119 tool requirements, 18 sector lanes, and 11 named authorities. These are derived counts, not fixed ceilings.
Executable ownership stays one-way: skills select, MCP exposes, the outer SDK routes, the internal SDK executes, ENV selects, UOP governs, tools perform bounded work, hooks emit receipts, and the owning authority validates effects.
Any missing identity, schema, grant, capability, dependency, receipt, or authority proof must fail closed at its owning phase; a later green check cannot retroactively authorize the skipped boundary.
A changed route refreshes every dependent schema, manifest, generator, test, diagram, and documentation reference; the superseded executable route is directly purged in the same Delta.
Tests, Git, CI, installation, restart, deployment, discussion, or a rendered page never imply Project HIL, Learning HIL, Goal completion, or pointer movement.

This page is a Git-tracked documentation projection. Executable source, SQLite authorities, installed-runtime receipts, and explicit human gates remain the governing evidence.