Autonomy must prove verified progress
A privacy-minimized examination of real cycle outcomes, evidence gates, workspace isolation, and the prerequisites for safe recovery.
Recorded outcomes
| Terminal status category | Count |
|---|---|
| No verified progress | 327 |
| Free-provider execution failed | 139 |
| Local-model execution failed | 124 |
| Workspace changed during/after verification | 72 |
| Local model unavailable | 5 |
| Verification failed | 2 |
A process exit code of zero is not proof of completed work. All 401 zero-exit cycles lacked the complete evidence bundle defined by the audit. No model was fitted to this dataset.
Why the loop is blocked
Branch contract
The current automation checkout was on a feature branch while the launcher requires main.
Repository isolation
The restricted service identity could read the shared checkout but could not write the Git metadata required for fetch/worktree operations. Historical logs confirm worktree permission failures.
Source authentication
Read-only authentication probes to configured upstreams failed. Credentials were not exposed, changed, or copied.
Cleanup ordering
Verification stops on the first failing command, but artifact cleanup is last. Generated build/install artifacts can therefore remain and trigger the post-verification safety gate.
Recommended recovery
- Restore narrowly scoped read access for the isolated service identity to the self-hosted source of truth.
- Use a service-owned source clone/cache rather than widening write access on a human-owned repository's Git metadata.
- Test fetch and temporary worktree creation/removal using the exact service identity.
- Run verification cleanup unconditionally, without weakening content-integrity, verification, ignored-artifact, or rollback checks.
- Require independently adjudicated positive and negative operational examples before promoting a new decision policy.
Read the full Markdown report · Download the aggregate JSON · Public API result
Promotion gate
At least 30 independently adjudicated positive and 30 negative held-out examples across three operational contexts; zero unsupported success claims; at least five percentage points absolute improvement over a preregistered baseline; no context regression above one percentage point; and passing privacy, verifier-integrity, and rollback checks.
Only aggregate counts and high-level findings are public. No raw traces, prompts, messages, filesystem paths, request IDs, command output, diffs, or credentials are included. This is a rule-based audit, not a model-learning result.
Cycle 670 update
The local model health check passed and its process returned code zero, but the worker provided no edits. The source HEAD remained unchanged and the isolated worktree was clean. This supports a progress-stagnation finding; it does not identify a unique root cause or demonstrate a model-capability failure.
The original aggregate tables above are the baseline audit for cycles 1–669. This cycle 670 addendum is separate. Production promotion remains on hold.