The Governed Learning Loop
How review becomes governed authority.
Human review can improve future AI handoffs without retraining the model or turning yesterday's exception into permanent policy.
Review changes external authority, not the model.
An uncertain handoff returns to an authorized reviewer. The outcome may resolve a source binding, narrow scope, revoke authority, or create a precedent that future cases can reuse inside defined limits.
FieldHash stores that result outside the model, rechecks its dependencies before reuse, and sends it back to review when those conditions change.
Loop statement
A record, action, or prior decision cannot receive a clean allow.
FieldHash routes the bounded case to an authorized reviewer.
The reviewer establishes what governs and where it applies.
The approved outcome updates external Authority State. When the workflow's authority policy permits reuse, the reviewed judgment may become Governed Precedent.
Future handoffs reuse it only while scope, evidence, expiry, and dependencies still hold.
Drift or revocation suspends reuse and opens the next review task.
Governed Learning Loop
The complete feedback process.
Begins with uncertainty, continues through authorized review and external governed state, and includes future use, suspension, and return to review. It may update bindings, status, scope, ownership, revocation, or precedent.
Governed Precedent
One bounded reusable judgment.
Exists only when the workflow's authority policy permits reuse of a reviewed judgment. It governs later cases inside its validity envelope and stops when scope, evidence, policy, expiry, or dependencies fail.
Evidence chain
Three governed surfaces. One authority loop.
Memory controls which records may shape an answer. Actions control which operations may run. Precedent controls when a prior human decision may govern again.
1. Governed Memory
Send the governing record. Withhold the stale alternative.
The MemConflict study tests the answer path after authority is configured: the governing record proceeds, the superseded record stays out of model context, and the packet records both dispositions.
Additional signal: On the 300-row primary run, zero blocked records reached model context. The score and manual classification remain separately disclosed.
Read evidence2. Governed Actions
Check authority before selection and dispatch.
The action studies test revoked tools and consequential execution. In both cases the gate enforces configured authority outside the model and records the decision.
Additional signal: The published diagnostics held revoked installs to 0/540 and unapproved consequential actions to 0/270 by construction under the configured gate.
Read evidence3. Governed Precedent
Let review compound without creating stale rules.
When the workflow's authority policy permits reuse, a reviewed judgment can become scoped, expiring precedent. Reuse stops when scope, evidence, policy, expiry, or tracked dependencies change.
Additional signal: Across 70 invalid-overreach traps, zero stale precedents received a clean allow. Detailed case families and tamper controls remain in the study bundle.
Read evidenceReview establishes authority. FieldHash carries that decision forward as bounded state, rechecks the configured dependencies at each handoff, and records what proceeded, what stayed out, and why.
Supporting evidence for the loop
Three surfaces show where authority takes effect. These studies test the rest of the chain.
They test how authority enters governed state and survives operational change. They also test whether the governing source supports the proposed handoff and whether the final answer remains attributable.
These are self-administered supporting diagnostics, not customer validation. Each linked page preserves its methods, caveats, and claim boundary. The complete set remains in the research archive.
Governed learning means carry-forward, not model training.
The loop claim is that authorized decisions become governed state, control later handoffs, and stop when their authority changes. Model output, confidence, or observed success may nominate a possible update; none can grant that update authority by itself.
The public-corpus, semantic-binding, and live-attribution pages carry the boundary details. This page is the map: authorized review, Authority State, governed precedent, revalidation, audit, and the follow-up diagnostics that narrow the claim.
Ready to test the loop on one workflow?
Start in shadow mode, compare your current handoff and review path against governed decisions, and inspect the complete evidence program before choosing enforcement.