Evidence-Gated Rule Promotion
Promote a local learned behavior only after checking recurrence, scope and counter-evidence.
These examples and illustrative results are independently authored teaching materials, not measured model results.
Use case
Two independent services support boundary validation while a React folder convention fits one project only. Promotion needs compatible triggers/actions/scope/corrections, not session confidence.
Mechanism
Retain observations/results and establish independence/context. Compare existing rules, framework limits and corrections. Propose minimal transferable action with projects/exceptions; contradictions/locality downgrade or defer. Scores are workflow signals and governed promotion needs authority, without unrequested memory writes.
Bad example
Promote one React success globally, call confidence 0.8 an 80% correctness probability and ignore corrections.
Good example
Cite both services’ compatible observations and inspect counterexamples/existing coverage before wider proposal. Keep React structure local and contradictions pending; source 0.8 is configuration, not calibration. Maintenance requires actual authorization.
Why the change matters
Cross-context evidence supports transfer more than one experience, but constraints/corrections define applicability. Retained observations prevent abstract rules outgrowing their conditions.
Observable expectation
Teaching copied reports are not independent observations. A correction against duplicate internal validation narrows the rule to boundaries instead of validate everywhere. Projects/triggers/results remain inspectable.
Limits
Recurrence is not causality/truth and no correction is not approval. Confidence is not probability and project policies may override. No memories/global rules were modified.