Data

Public evidence catalogue

Every accepted run below, with its structured metadata and downloadable raw evidence. Before publishing, raw files are passed through a fixed set of redactions: local home-directory and Windows drive paths (reduced to the filename), IPv4 addresses, PEM private-key blocks and SSH public keys. That is the whole list — it is not a general secret scanner. Every other value, including all measurements and hashes, is published exactly as captured.

What this evidence covers

Evidence level: E1 Local observation: a performance run on one host per system, with incomplete host control. Not independently reproduced.
Measured on: 27–28 September 2026
Tested configuration: standard-ai-v0.1 · llama.cpp build 11146 · context depth 8192 · 512 prompt / 128 generated tokens · 5 timing repetitions in one session (repetitions, not independent replications)
What was measured: Prompt-processing and generation throughput, in tokens per second.
What PASS means: The run completed without a failure classification (model loaded, every layer offloaded to the targeted accelerator, no CPU fallback, timeout or crash) and was reviewed into the accepted registry. It is not a verdict on answer quality or usability.
Not measured: Answer quality (semantic capability)
Useful context length
Time to first token
Interactive responsiveness
Thermal behaviour
Power draw
Largest model that fits
Multitasking alongside other work

Apple M1 Pro (32 GB unified memory)

Result ID20260927T165914Z-d4d62449

Timestamp2026-09-27T16:59:14.790462+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorApple M1 Pro

Benchmark target (recorded selection)BLAS,MTL

Modelllama 8B Q4_0

Model SHA-2566a0746a1ec1aef3e7ec53868f220ff6e389f6f8ef87a01d77c96807de94ca2aa

BackendBLAS,MTL

Prompt processing190.39 tok/s

Generation27.92 tok/s

HostmacOS / Apple M1 Pro

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

Result ID20260927T170032Z-ef58c371

Timestamp2026-09-27T17:00:32.230102+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorApple M1 Pro

Benchmark target (recorded selection)BLAS,MTL

Modelqwen3 14B Q4_K - Medium

Model SHA-256a8cc1361f3145dc01f6d77c6c82c9116b9ffe3c97b34716fe20418455876c40e

BackendBLAS,MTL

Prompt processing97.09 tok/s

Generation11.73 tok/s

HostmacOS / Apple M1 Pro

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

AMD Radeon RX 9070 (16 GB dedicated VRAM)

Result ID20260927T182043Z-d8f20dab

Timestamp2026-09-27T18:20:43.708094+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorAMD Radeon(TM) Graphics

Benchmark target (recorded selection)AMD Radeon RX 9070 (targeted via -dev Vulkan1; Vulkan)

Modelllama 8B Q4_0

Model SHA-2566a0746a1ec1aef3e7ec53868f220ff6e389f6f8ef87a01d77c96807de94ca2aa

BackendVulkan

Prompt processing2511.81 tok/s

Generation102.31 tok/s

HostMicrosoft Windows 11 Pro / AMD Ryzen 7 9800X3D 8-Core Processor

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

Result ID20260927T182108Z-d90c9293

Timestamp2026-09-27T18:21:08.188128+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorAMD Radeon(TM) Graphics

Benchmark target (recorded selection)AMD Radeon RX 9070 (targeted via -dev Vulkan1; Vulkan)

Modelqwen3 14B Q4_K - Medium

Model SHA-256a8cc1361f3145dc01f6d77c6c82c9116b9ffe3c97b34716fe20418455876c40e

BackendVulkan

Prompt processing1177.22 tok/s

Generation56.38 tok/s

HostMicrosoft Windows 11 Pro / AMD Ryzen 7 9800X3D 8-Core Processor

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

AMD Radeon RX 7800 XT (16 GB dedicated VRAM)

Result ID20260928T134138Z-87cbbded

Timestamp2026-09-28T13:41:38.483282+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorAMD Radeon RX 7800 XT

Benchmark target (recorded selection)Vulkan

Modelllama 8B Q4_0

Model SHA-2566a0746a1ec1aef3e7ec53868f220ff6e389f6f8ef87a01d77c96807de94ca2aa

BackendVulkan

Prompt processing763.53 tok/s

Generation74.81 tok/s

HostMicrosoft Windows 11 Pro / AMD Ryzen 7 3700X 8-Core Processor

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

Result ID20260928T134214Z-4c21f67d

Timestamp2026-09-28T13:42:14.102398+00:00

Methodologystandard-ai-v0.1

StatusPASS

First enumerated acceleratorAMD Radeon RX 7800 XT

Benchmark target (recorded selection)Vulkan

Modelqwen3 14B Q4_K - Medium

Model SHA-256a8cc1361f3145dc01f6d77c6c82c9116b9ffe3c97b34716fe20418455876c40e

BackendVulkan

Prompt processing417.30 tok/s

Generation43.51 tok/s

HostMicrosoft Windows 11 Pro / AMD Ryzen 7 3700X 8-Core Processor

result.json ↓manifest.json ↓system.json ↓llama-bench.json ↓stdout.log ↓stderr.log ↓

Model file paths in raw evidence are redacted to <local-path-redacted>; the filename itself is preserved. SHA-256 hashes above are computed from the model file's own bytes and are unaffected by that redaction.