Concepts / architecture

Execution is not
evidence.

Forge separates the plane that runs a worker from the plane that decides whether the outcome is real, reusable, and safe to carry forward.

Product illustration. It describes the contract; it is not a live AO session view.

The flow

01 / executeAO or another harness

Runs the worker in a project/worktree context. Forge does not replace the harness.

02 / verifyFrozen or independent checks

Tests, changed-file evidence, and artifact inspection decide whether the result passes.

03 / repairBounded recovery

A failure can receive one evidence-based repair attempt; no-op and hidden-blocked sessions are stopped.

04 / reflectCandidate skill

A model may summarize the failure into a structured candidate. It cannot promote itself.

05 / gateG1 / G2 / G3

Applicability, A/B benefit, and held-out no-regression evidence are required for reuse.

observedBacked by execution and readback

Example: C0 baseline failed, bounded repair passed, and the local regression suite passed.

candidateExists, but not promoted

Example: GPT-5 Nano returned a structured skill candidate with a bounded scope.

unverifiedRequired proof is still open

Example: reliable autonomous AO spawn/completion and second-harness transfer.

AO: health/readiness, catalog reads, and isolated worktrees are observed; final artifact-producing completion remains unverified.
Hermes: a reflection sidecar and proposal source, not an AO worker or promotion authority.
Neatlogs: the real reflection/submission workflow has a verified trace readback; this does not prove every worker path is traced.
Supermemory: future external skill/context adapter; local registry and ledger remain authoritative now.