Verification principles
Transparency
Wins and misses are reported together each month.
Reproducibility
Repro commands and artifact paths are retained for audit.
Known limits
False positives and misses can occur and are disclosed upfront.
Current evidence snapshots
Readiness `l1_inverse_decoder_commercialization_checklist_v1.json` → READY_FOR_BROAD_ROLLOUT (3/3 criteria passed).
Hard-case gain `swap_typo` stress mode with objective v4 improved exact/recovery by +0.0667.
Scope boundary Meaning-centric workloads first. 100% literal-fidelity workloads should use literal-preserve lanes.
Operational safety controls
- Improved path is default-on with immediate rollback switch.
- Track A and Track B are separated for governance and promotion control.
- No direct commercialization claim without reproducible artifacts.
Example reproducibility command
bash scripts/deploy/linux/check_a_codeai_public_routes.sh
Auto-refreshed evidence snapshot
Live health check
loading...
Commercial readiness (auto)
loading...
Hard-case gain (auto)
loading...
Refresh policy
This block is populated from `/health` and `/evidence/latest.json` on page load with `cache: no-store`.
Open validation pack
Request a pilot pre-check pack including route checks, benchmark reading guide, and operation boundary checklist.