Adaptation is intentionally narrower than general training. Version 1 can update only the final linear head, and only when both mode=adaptive and learning_mode=verified-head.
Required feedback
| Field | Purpose |
|---|---|
feedback_id | Stable audit identity. |
evaluation_fingerprint | Binds feedback to the exact evaluated model, Monk policy, features, and scores. |
targets | One or more known output labels with bounded integer target scores. |
verifier | Host-controlled identity or verification method. |
evidence_digest | SHA-256 of the host's supporting evidence. |
verified=true | Explicitly enters the proposal path; absence fails closed. |
Bounded update
The error is the target minus current score. Weight and bias deltas use the configured parts-per-million learning rate, deterministic signed division, and per-parameter step caps. Updated values saturate to the attachment limit.
Proposal result
The result contains the current and proposed attachment fingerprints, before and after evaluation fingerprints, normalized feedback, every changed parameter, committed=false, and authority=proposal-only. The proposed attachment increments its revision and names its parent fingerprint.
Host responsibility: Loom verifies structural binding, not the truth of the external evidence. The host decides who or what may set
verified, retains the supporting evidence, reviews the delta, and uses the ordinary build/verify/commit path if accepted.