Completion audit
The user authorized keeping partial support experimental and persistent failures or missing prerequisites unstable rather than forcing all adapters to pass. Under that scope, qualification is complete; universal 29/29 compatibility is not claimed.
| Requirement | Verified evidence |
|---|---|
| Existing 29 adapters, four profiles, two fixed tasks | matrix.json has exactly 116 unique cells, 29 adapters, four providers and two completed task attempts per cell. No pending cells. |
| Pinned provider configuration | manifest.json and per-attempt qualification/manifest files record OpenAI and Anthropic dated model IDs, HF's explicit Together model route, and Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. HF route is not an immutable weights revision. |
| Exact TiTO and real optimizer | 21 current vLLM cells have matching optimizer proofs: 99 current rows. Jobs 79350, 79704 and 79761 completed 85, 8 and 18 diagnostic steps respectively, all finite/nonzero-gradient steps. Prior captures remain historical and do not qualify newer ones. |
| All outcomes recorded honestly | 21 OpenAI, 20 Anthropic, 19 HF and 21 vLLM passes. The other 35 cells are failed with artifacts/prerequisites. Stable 14, experimental 9, unstable 6. |
| Deterministic regression | harbor-regression-final.log: 567 passed, 2 skipped. Actual Anthropic SDK stream replay passed separately. Evidence helper tests passed after task identity checks. Ruff, usort, formatter and git diff checks pass. |
| Current source integrity | snapshots/release-final-v1/sha256.json verifies 1,483 files. Worktree source matches. The only change from the optimizer's release-candidate source is AST-equivalent Ruff formatting, proven in format-equivalence.json. |
| Gradio and training contract | Core export is authoritative; eval has no training export; tier defaults, experimental opt-in, per-rollout ACP/NeMo profiles and callback state have regression coverage. Generic registry is not mutated. |
| Repeatable targeted tests | tools/run_provider_cohort.py accepts paths, task indices, harness selection and new immutable attempt names. Fixed-task names and file hashes are checked before matrix promotion. Per-attempt snapshots are retained. |
| Versioned report | final-v1 contains the frozen matrix, provider manifest, limitations, checks, audit and rendered report with artifact hashes. Raw capture evidence remains under the stated evidence root. |
| Isolation and cleanup | Existing training job 79083 remained running. Existing eval 79305 completed 0:0. Qualification GPU jobs used hopper-dev; no hopper-extra. Sandbox audit found zero owned remnants. Only qualification job 79176 was retired after its consumers finished. No pushes or shared-environment installs. |
Explicit limits
These are compatibility smoke tests, not benchmark pass@1 scores or production-scale guarantees. Optimizer replay uses diagnostic advantage +1; weight synchronization, reward-normalized learning, long-run stability and fair multi-row rollout weighting are not validated by this qualification. ACP and NeMo evidence applies only to the explicitly named workflow profiles. Native ATIF token-count omissions are reported, not fabricated. Missing vendor credentials/apps and failed routes remain excluded from stable defaults. Existing live training services were deliberately not redeployed with these local changes.