feat(self): add bounded improvement learning loop
This commit is contained in:
@@ -2,6 +2,8 @@
|
||||
|
||||
## Gen2 M6 checkpoint — 2026-08-30
|
||||
|
||||
- `GEN2-EVOLVE-004 = MILESTONE_ACCEPTED`: bounded SELF learning signals, evidence-backed attribution, contextual weakness/competence, ImprovementNeed, independently evaluated candidate, benchmark and expected-vs-actual Learning Ledger primitives are implemented with durable SELF persistence in migration 055. Evidence: `docs/evidence/gen2-evolve-004-milestone-2026-08-30.md`. No autonomous promotion, production deployment, new owner, UI work, V1 retirement or GEN2-EVOLVE-005 was started.
|
||||
|
||||
- `GEN2-EVOLVE-003 = MILESTONE_ACCEPTED`: Experience evidence now enters an evidence-governed Candidate Skill → independent validation → Active Skill → bounded Candidate Reflex lifecycle. Reflex divergence/revalidation records SELF proposals and Learning Ledger evidence; it cannot autonomously rewrite or promote production. Evidence: `docs/evidence/gen2-evolve-003-milestone-2026-08-30.md`. UI and V1 compatibility remain unchanged.
|
||||
|
||||
- `GEN2-EVOLVE-002 = MILESTONE_ACCEPTED`: one bounded typed cognitive memory-continuation round is available through Core and the existing retrieval fabric. Evidence: `docs/evidence/gen2-evolve-002-milestone-2026-08-30.md`. UI and V1 compatibility remain unchanged.
|
||||
|
||||
@@ -0,0 +1,28 @@
|
||||
# GEN2-EVOLVE-004 milestone evidence
|
||||
|
||||
Status: MILESTONE_ACCEPTED for the bounded SELF evidence/proposal domain.
|
||||
|
||||
## Delivered
|
||||
|
||||
- Typed learning signals cover verified failures, corrections, repeated mistakes, Reflex divergence, Skill fallback, escalation, outcome mismatch, retry/fallback, stale dependencies, regret and better solutions.
|
||||
- Attribution returns a bounded evidence-backed `HYPOTHESIS`; provider outage is not converted into a competence loss.
|
||||
- Weakness patterns require two matching incidents in the same Principal/task/Scope and retain evidence, frequency, severity, affected component and confidence.
|
||||
- Competence assessment is contextual and requires evidence; SELF cannot self-declare trust.
|
||||
- ImprovementNeed and Candidate contain hypothesis, expected effect, evaluation, must-not-worsen, rollback, cost and independent evaluator fields.
|
||||
- Benchmark cases reference evidence rather than copying conversations. Learning Ledger appends expected and actual evaluation references and decision.
|
||||
- Candidate evaluation rejects latency/quality regression and creator/evaluator identity reuse. New model qualification creates a review candidate only.
|
||||
- PostgreSQL migration `055` adds SELF-owned durable signal, weakness, candidate and benchmark tables under existing `self_role`; no new role, owner or cross-owner relation is created.
|
||||
|
||||
## Checks
|
||||
|
||||
- `go test ./contracts/self ./modules/self/domain ./adapters/postgres ./cmd/core-runtime`
|
||||
- `go vet ./contracts/self ./modules/self/domain ./adapters/postgres ./cmd/core-runtime`
|
||||
- `go run ./cmd/devctl preflight -task dev/tasks/GEN2-EVOLVE-004-self-improvement-loop-weakness-diagnosis-and-governed-learning.json`
|
||||
- `go run ./cmd/devctl check-diff -task dev/tasks/GEN2-EVOLVE-004-self-improvement-loop-weakness-diagnosis-and-governed-learning.json`
|
||||
- `go run ./cmd/devctl/independent`
|
||||
- `go run ./cmd/devctl check-architecture`
|
||||
- `git diff --check`
|
||||
|
||||
## Boundaries
|
||||
|
||||
No autonomous code generation, merge, deployment, production promotion, new SELF owner, Skill/Reflex mutation, UI work, V1 retirement or Gen2-EVOLVE-005 was performed. Existing Work governed-evolution gates remain the promotion authority. Runtime recorder wiring through `cmd/core-runtime/main.go` was not added because it is outside the accepted packet scope; the durable repository and typed domain path are ready for a separately authorized integration task.
|
||||
Reference in New Issue
Block a user