| # Release validation |
|
|
| ## November evidence refresh |
|
|
| This documentation release aligns the model card and evidence guide with the |
| audited SAR manuscript. Executables, client, dependency requirements, IBM |
| weight/source pin and the greedy-plus-one-sample strict-0.05 policy are unchanged |
| from the previous November bundle. The paper's new research controls use separate |
| frozen runners; they are not extra models or policies in this download. |
|
|
| `validation/maintenance-20261106.json` records the new offline package/source |
| checks and comparison with the previous anonymous artifact. |
| `validation/maintenance-20261102.json` remains intact as historical evidence. |
| No real-model inference, live Jev call or new GPU allocation belongs to this |
| package-maintenance check. Optional inference tests are reported as skipped |
| where their dependencies are absent, not treated as numerical validation. |
|
|
| The [source publication receipt](https://github.com/OVRLab/jev-guided-decoding/blob/main/releases/granite-jev-1b-sampling/publication.json) |
| records the final combined-source checks and anonymous exact-revision download. |
| The [installation guide](https://huggingface.co/OVRLab/granite-jev-1b-sampling/blob/main/INSTALL.md) |
| and publication metadata remain outside the hashed bundle to avoid circular |
| revision references. Inference-library pins and their disclosed security |
| boundaries are unchanged; see [SECURITY.md](SECURITY.md). |
|
|
| ## Preserved real-model and provider evidence |
|
|
| `validation/release-smoke.json` and `validation/release-e2e.json` preserve the |
| October 2026 Azure CPU integration checks. They are historical receipts, not |
| new tests of this maintenance environment: |
|
|
| - IBM Granite 4.0–1B FP32 greedy/sample token parity against the frozen R51 |
| runtime passed on an authored arithmetic case (two tokens per mode), with |
| unchanged weights and matched seed/payload mappings. |
| - An authored full solve generated 43 native and 45 sampled tokens, called Jev |
| 1.13.0, and returned the exact chosen native token sequence; the provider |
| reported 638 input and 40 output tokens. |
| - The original offline tests covered strict margin, reference isolation, token |
| ownership, invalid judgments, credentials, cache isolation and failure records. |
| Subsequent regressions added interruption, durable journaling, hooks and usage |
| accounting. See the existing source tests for the full contract. |
|
|
| The [previous validation record](https://huggingface.co/OVRLab/granite-jev-1b-sampling/blob/5f85775870e1dabe060698d5c5f1506a138f6b06/VALIDATION.md) |
| preserves the detailed history. Neither those small integration cases nor the |
| current unit tests establish answer quality, current provider availability, |
| GPU/cross-device equality or deployment throughput. Scientific evidence is |
| separate and summarized in [RESULTS.md](RESULTS.md). |
|
|
