Only rows backed by retained certification reports belong here. Empty coverage means unverified, not failed.
All rows below were produced by scripts/validate_windows.ps1 with mock mode disabled on a single physical machine. Requested and actual devices were recorded independently; AUTO results are reported separately and are never used to infer a direct GPU or NPU claim. The sanitized reports are published under certification/0.6.0/; raw reports, including server logs, remain local under the ignored certification/results/ path. Verify any report against certification/0.6.0/SHA256SUMS.txt, and see certification/0.6.0/README.md for full scope, meanings, and limitations.
Machine: Windows 11 Home build 26200, x64, Intel Core Ultra 9 185H, Intel Arc iGPU (adapter driver 32.0.101.8425), Intel AI Boost NPU (driver version not captured in the retained reports), 31.7 GB RAM.
The reports record server version=0.5.0: the certification sessions ran on 2026-07-24 at ~09:17–09:32 UTC, immediately before the commit that bumped app/version.py to 0.6.0 (~09:53 UTC the same day). The qualified OpenVINO 2026.2.1 runtime and all exercised code paths are identical between that pre-bump tree and the 0.6.0 release.
| App version | Windows | Arch | Processor | OpenVINO | GenAI | Device | Driver | Model | Precision | Requested | Actual | Load | Benchmark | Report | Date |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | CPU | (integrated, no discrete driver) | tinyllama-1.1b-chat-fp16 | FP16 | CPU | CPU | PASS | PASS | api-cpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | GPU (Arc iGPU) | 32.0.101.8425 | tinyllama-1.1b-chat-fp16 | FP16 | GPU | GPU | PASS | PASS | api-gpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | NPU (AI Boost) | not captured in report | tinyllama-1.1b-chat-fp16 | FP16 | NPU | NPU | PASS | PASS | api-npu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | CPU | (integrated, no discrete driver) | smollm2-135m-fp16 | FP16 | CPU | CPU | PASS | PASS | api-cpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | GPU (Arc iGPU) | 32.0.101.8425 | smollm2-135m-fp16 | FP16 | GPU | GPU | PASS | PASS | api-gpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | NPU (AI Boost) | not captured in report | smollm2-135m-fp16 | FP16 | NPU | NPU | PASS | PASS | api-npu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | CPU | (integrated, no discrete driver) | qwen2.5-1.5b-fp16 | FP16 | CPU | CPU | PASS | PASS | api-cpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | GPU (Arc iGPU) | 32.0.101.8425 | qwen2.5-1.5b-fp16 | FP16 | GPU | GPU | PASS | PASS | api-gpu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | NPU (AI Boost) | not captured in report | qwen2.5-1.5b-fp16 | FP16 | NPU | NPU | PASS | PASS | api-npu.md | 2026-07-24 |
| 0.6.0 | 11 build 26200 | x64 | Ultra 9 185H | 2026.2.1 | 2026.2.1.0 | CPU | (integrated, no discrete driver) | bge-small-en-v1.5 (embeddings) | FP16 | CPU | CPU | PASS | n/a (384-dim) | api-cpu.md | 2026-07-24 |
AUTO was also certified PASS for tinyllama-1.1b-chat-fp16 (api-auto.md), smollm2-135m-fp16 (api-auto.md), and qwen2.5-1.5b-fp16 (api-auto.md) in the same sessions. AUTO selects an OpenVINO device at runtime; those results validate the routing path and are intentionally not entered as direct CPU/GPU/NPU claims.
The smollm2-135m-fp16 rows record a model-dependent WARN on the tool/structured-output check (the request was accepted but the small model did not demonstrate both optional outputs). A WARN is not a failure. The tinyllama and qwen2.5-1.5b rows have no warnings.
The Benchmark column records whether the benchmark endpoint ran successfully on this machine; it is a functional check, not a throughput measurement, and is not a general speed guarantee. Only FP16 was certified; INT4/INT8 and contexts longer than the short validation prompts remain unverified.
These retained 0.6.0 reports predate the deterministic context-depth check and therefore
do not qualify any model_library_manifest.json certification record. Current harness
schema 2 reports must include a passing context_depth result with the exact requested
prompt-token depth, actual runtime device, and generated-token count. A manifest record
also requires the retained report path and its SHA-256. Until all of those facts exist,
the corresponding certification array and max_tested_context remain empty/zero.
The first schema 2 result is tinyllama-1.1b-chat-fp16 on CPU: requested and actual
device CPU, exact prompt depth 1,536 tokens, and 3 generated tokens. See
0.6.1/windows-certification-20260724-144444/windows-certification.json
(41590beb5ce497b37bf44189a39bc0a9fe95696c24a107b33cfeb545778f985b).
This context-capacity check makes no throughput claim; every other model/device
combination remains ineligible for the bundled manifest.
The issue #17 runs add deterministic context-boundary and compiled-cache restart checks. Every row below is backed by a sanitized schema 2 report. Context is exact prompt-token depth; generated tokens are listed separately. These checks make no throughput claim.
| Model | Precision | Requested | Actual | API | Context boundary | Cache restart | Overall | Report | SHA-256 |
|---|---|---|---|---|---|---|---|---|---|
| qwen2.5-3b-fp16 | FP16 | CPU | CPU | 12 PASS, 2 SKIP | 3,072 + 1 generated; 3,073 rejected | PASS | PASS | report | 6e01022c06ff309d4875fca75fd709c0ca3885ad7723862df443b9aeeabcbf10 |
| qwen2.5-3b-fp16 | FP16 | GPU | GPU | 12 PASS, 2 SKIP | 3,072 + 1 generated; 3,073 rejected | PASS | PASS | report | 27cadbef8a385b1014cfea075806a38d1b407769aa76cbdb1babff64859960db |
| qwen2.5-3b-fp16 | FP16 | NPU | not captured | 11 PASS, 2 SKIP, 1 FAIL | not run | not run | FAIL | report | d707c944a51d61dad96152e4a886623b258878c33c84ece52d771e028f980007 |
| qwen2.5-3b-fp16 | FP16 | AUTO | unknown | 12 PASS, 2 SKIP | actual device unknown | FAIL | FAIL | report | 5708b4b84b4eaf54a317f2d6f89fb3a9c88e1ef7b3fb7c2cbaf6572b5c3c3b77 |
| tinyllama-1.1b-chat-int4 | INT4 | CPU | CPU | 12 PASS, 2 SKIP | 1,536 + 3 generated; 1,537 rejected | PASS | PASS | report | 1cccdf519c008d91e8b0278237c52285c3a0e88b8d0386f8bbf7cfe3fa63d69f |
| tinyllama-1.1b-chat-int4 | INT4 | GPU | GPU | 12 PASS, 2 SKIP | 1,536 + 3 generated; 1,537 rejected | PASS | PASS | report | d46ed0d52e9b12a312b5257b2eaa74f2f6d50d395089067d6df5ca051f87ca23 |
| tinyllama-1.1b-chat-int4 | INT4 | NPU | not captured | 3 PASS, 2 SKIP, 9 FAIL | not run | not run | FAIL | report | ce59c1b457e217a43830fa05cb8a6113a4637d46b7b0f9ad65b8b6d80b73d499 |
| tinyllama-1.1b-chat-int4 | INT4 | AUTO | unknown | 12 PASS, 2 SKIP | actual device unknown | FAIL | FAIL | report | 5ef79ae3895ba67e02caf527111e6b6c6664ca908ca87dbb3380179d3198b664 |
Only the four direct CPU/GPU PASS combinations populate the bundled manifest. NPU remains empty for both models. AUTO is reported independently and never infers a direct-device claim. The conversion cancellation/retry evidence and package-footprint details are in the bundle README; verify all files against SHA256SUMS.txt.
Each report below is a clickable repository-relative link with its SHA-256. The complete list, including the JSON reports and this bundle's README, is in certification/0.6.0/SHA256SUMS.txt.
| Report | SHA-256 |
|---|---|
| 091712/windows-certification.md | 08efcc998478c50d47ea662ee3b5f6041a4a4491c48114c395cafe7420290f4b |
| 091712/api-cpu.md | d2f83524b9ed6c9477291dcdcad6cb0f574a3e73d583cb0709c4fa4007fd07cc |
| 091712/api-gpu.md | 933793bc68d4139c6b335eed9f79db86333a8b7f5ceb75440aee80e3cede4e03 |
| 091712/api-npu.md | f2721fadd6096a0e8b5bd2f613cb1e7ac7aae42cf78a757abf1ef1f6fa680035 |
| 091712/api-auto.md | 95cb6b75f093bbb1cb1b525b13dd8e5a5560f3a97e0098b7bef7df7fd5dd3675 |
| 092800/windows-certification.md | 8324bdea0ec7ed2624a1df2de1b03616f637bc41e30bf2676bbfe71f0e5d4822 |
| 092800/api-cpu.md | 7e611efbbed739494e8890125e54ecf395336d332b5bffbef8c9e73c8abae7bf |
| 092800/api-gpu.md | 6e7132d8d8a49d8e63656067ee13e2f1028a627979ad2142a3d0c5ad97169ae0 |
| 092800/api-npu.md | e1cc6cd3547ea3b68e7903420909095a171075900cfb82400cad76210ef0d3d9 |
| 092800/api-auto.md | 74eeb0936c1668dd9df0b7efdcaf41994d0cfce22b29bfbbf03c924c32578155 |
| 092853/windows-certification.md | 6c13c0d389d5a6e9f27a3c231a77ce30a24907bcf607acf30bd4fa7983d2f25b |
| 092853/api-cpu.md | 097269e633d2a93298f74f64df59417c9e5842c0637a552aa665e5ffa5711b5f |
| 092853/api-gpu.md | 63bf0939d62b52c4bfde7aeffa9c835487fd0ef3bb8572bc62573b60eaecafe7 |
| 092853/api-npu.md | 9faebd995378e02dd6d9368f638cd471be58ea9683c88bd7722d20a40ee69766 |
| 092853/api-auto.md | 881cf7842735b41e5333bd074e4a29c955f66b1d0b5c020a0fec547395fcb3d8 |
Mock contract results must not be entered as hardware certification. Add a row only after the Windows certification harness records the requested device, actual device, versions, driver, model, load result, benchmark result, report reference, and validation date.