# Maintainability canary FrontierLab v0.37 implements **Phase 1 maintainability: automatic detection, provenance capture and quarantine policy**. It does not perform autonomous repair. ## Trigger model The GitHub Actions workflow `.github/workflows/maintenance-canary.yml` runs: - manually (`workflow_dispatch`); - on relevant dependency / FrontierLab code changes; - daily at `02:17 UTC` (10:17 Asia/Taipei). The daily schedule matters because FrontierLab uses bounded dependency ranges. A new upstream package release can therefore change the resolved environment even when the repository itself has not changed. ## Canary sequence ```text install current dependency range ↓ run full regression suite ↓ capture exit code + pytest log ↓ record Python/package versions + requirements hashes ↓ compute dependency fingerprint ↓ HEALTHY or QUARANTINE_REQUIRED ↓ upload maintenance evidence ``` On failure the workflow still uploads its evidence and then exits non-zero. ## Governance boundary The health record explicitly fixes these controls: - `automatic_repair_enabled = false` - `repair_attempted = false` - `automatic_environment_rebuild_enabled = false` - `automatic_commit_enabled = false` - `automatic_merge_enabled = false` - `human_approval_required_before_promotion = true` A failing revision/environment therefore enters **QUARANTINE_REQUIRED** rather than being silently repaired or promoted. ## First executed canary Feature-branch GitHub Actions run: `41064072834` Artifact: `maintenance-canary-41064072834` (artifact ID `12980538164`) Recorded result: | Field | Value | | --- | --- | | Status | **HEALTHY** | | FrontierLab version | 0.36 at execution time | | Python | 3.11.17 | | Streamlit | 1.66.1 | | Requests | 2.34.2 | | OpenAI | 2.54.0 | | Pytest | 9.2.0 | | Regression tests | **361 passed** | | Dependency fingerprint | `ac61d837d7b2c80e5f0e6e3cb1fe6a0f5cb42705a52dba18fa2dee71f68d4b34` | | Automatic repair | No | | Promotion allowed | Yes, because regression tests passed | The artifact also preserves SHA-256 hashes of `requirements.txt` and `requirements-dev.txt` plus the exact source SHA. ## What this proves The canary proves that FrontierLab can automatically: 1. instantiate the current allowed dependency environment; 2. run regression tests; 3. detect a software-level pass/fail state; 4. record environment/dependency provenance; 5. preserve the test log and health report; 6. enforce a quarantine decision on failure. It does **not** prove: - autonomous repair; - scientific reproducibility; - scientific accuracy; - safety of dependency upgrades beyond the software test suite. ## Next maintainability phase The next research phase can compare: - manual maintenance; - automatic detection + human repair; - automatic detection + bounded repair proposal + human promotion approval. Any repair experiment should remain isolated and should never automatically merge into trusted main. ## Final v0.37 validation After the central version guard and canary trigger scope were synchronized, FrontierLab reran both the ordinary test workflow and the Maintenance Canary on commit `2a076c1d8b485cdc47e0352a2d5fea412753db0e`. - Maintenance Canary run: `41064836504` - Artifact: `maintenance-canary-41064836504` - Artifact ID: `12982122350` - FrontierLab version: **0.37** - Status: **HEALTHY** - Regression tests: **361 passed** - Dependency fingerprint: `ac61d837d7b2c80e5f0e6e3cb1fe6a0f5cb42705a52dba18fa2dee71f68d4b34` - Failure detected: **false** - Automatic repair attempted: **false** - Human approval before promotion: **required** - Promotion allowed by software canary: **true** The canary trigger scope now includes FrontierLab package code, `app.py`, scripts, tests, dependency files and the canary workflow itself, in addition to the daily schedule. ## Phase 2 advisory proposal The canary now also emits `maintenance_repair_proposal.json`. It compares the current environment with the human-controlled Nov 1 healthy baseline and returns one of: - `NO_REPAIR_NEEDED`; - `DRIFT_REVIEW_REQUIRED`; - `REPAIR_PROPOSAL_READY`. The first integrated Phase 2 canary was GitHub Actions run **41297646386**, artifact **13079788250**. It passed **367 software tests**, matched the accepted dependency fingerprint, detected no package/requirements/Python drift, and emitted: ```text status = NO_REPAIR_NEEDED classification = HEALTHY_STABLE automatic_repair_enabled = false automatic_code_edit_enabled = false automatic_dependency_edit_enabled = false automatic_commit_enabled = false automatic_merge_enabled = false automatic_deploy_enabled = false ``` See [Advisory maintenance repair proposals](MAINTENANCE_REPAIR_PROPOSAL.md).