Prove evidence-gated learning end to end

[Haz3-jolt]

Outcome

Add a bounded end-to-end Phase 11 suite proving that learned changes are visible, evidence-backed, scoped, reversible, regression-tracked, and unable to execute generated code without explicit approval.

Current repository evidence

• Instruction loading, Agent Skills, and prompt templates are implemented foundations.
• No learning ledger, insights pipeline, managed-rule writer, generated-artifact quarantine, learning notification capability, or Phase 11 integration suite exists.

Acceptance criteria

• Drive an explicit user correction through candidate classification, an automatic tier 1 global managed-block change, ledger append, attached-client notification, and exact rollback using public daemon and SDK operations.
• Prove repository, tool, web, model, imported, and summary evidence cannot promote a global rule, and prove project-derived output stays under the documented project .axl/ scope unless the user explicitly promotes it.
• Generate a skill and prompt template through their existing formats, and route an SDK workflow program, hook, and native extension through quarantine with no executable auto-activation.
• Show that an insights report is candidate input only, while an explicit user acceptance creates the user-originated evidence needed for promotion.
• Trigger a confidence-bearing regression, group co-activated changes, auto-revert only an automatically activated tier 1 rule, and propose manual-tier removal without automatic mutation.
• Prove disable and rollback remove prompt, UI, process, and background contributions and that no old session content is retrieved or injected without an explicit recall pull.

Verification

Use synthetic session logs, deterministic fake providers, fake clocks, temporary global and project homes, and sandbox fixture processes. Exercise public operations and inspect canonical events, ledger entries, artifacts, notifications, and prompt inputs. No tests were run during this source audit.

Coordination

This is an integration gate, not another learning implementation. It consumes the focused Phase 11 issues and the shared public extension, SDK, sandbox, and sidecar substrates.

Source

ROADMAP.md, Technical implementation roadmap, Phase 11 Exit gate, line 2266; Product vision §8 Learning loop.

Repository audit

Audited against Axl 88944a2. Phase 11. Source inspection, not an executed behavior test. Roadmap references: L2266.

Tracking key: axl-backlog:learning.prove-phase11-exit-gate.

Issue tracking

Axl issue mirror: #369 | GitHub #369

Related issues:

• #128 (Build disposable per-session sidecar indexes and aggregate statistics #128)
• #208 (Create the append-only learning ledger #208)
• #211 (Measure learning outcomes and revert regressing automatic rules #211)
• #217 (Quarantine and approve generated workflow programs, hooks, and extensions #217)