Contents
Top-level results documents
| File | Contents |
|---|---|
| model-survey.md | Capability-survey log: per-model entries (capture format, corridor checks), Tier 1/2 summaries, quant ladder table |
| capability-probes.md | Non-LLM probes: whisper STT (+RTF), Piper TTS, FLUX image generation, resolution ceiling, 50-image gallery, upscaler |
| experiments.md | Pre-registrations AND results for experiments E1-E16 (sawtooth, quant-pair concurrency, upscaler, energy, spec decode, co-residency, 84K probe, embeddings, eval suites) |
| eval-pilot.md | Quality-eval pilot matrix, prediction resolutions, grader-artifact documentation |
| flux-gallery-prompts.txt | The 50 gallery prompts (filename|prompt), reusable |
| INDEX.md | this file |
Related (outside results/): docs/PHASE-A-LOG.md (Phase A evidence log), docs/FINDINGS.md (findings register F1-F21), reports/canonical-benchmark-report.md (compiled source of truth), manifests/MANIFEST.md (artefact provenance: repo, revision, size, sha256 for all 27 model files).
results/phase-a-raw/ - Phase A + Steps 8-9 raw outputs (47 files)
| Pattern | Contents | Discussed in |
|---|---|---|
| bench-step5-gate.md, bench-step5-longctx.md | gpt-oss-120b house reference baseline (d0, d8192) | PHASE-A-LOG step 5 |
| bench-step6-qwen3-30b-a3b-2507*.md | Workhorse matrix row (d0 + longctx) | PHASE-A-LOG step 6 |
| bench-step6-qwen3-14b.md | Dense 14B matrix row | PHASE-A-LOG step 6 |
| bench-step6-glm45air*.md | GLM-4.5-Air (split invocations, F6) | PHASE-A-LOG step 6 |
| bench-step7-gptoss20b.md | gpt-oss-20b row | PHASE-A-LOG step 7 |
| bench-step7-qwen36-35b.md | Qwen3.6-35B hybrid row (F7) | PHASE-A-LOG step 7 |
| bench-step7-glm47flash.md / -mxfp4.md | GLM-4.7-Flash quant-isolation pair (F11) | PHASE-A-LOG step 7 |
| bench-step8-depths-*.md (6 files) | Six-model depth curves, -d 0..32768 (F13/F14) | PHASE-A-LOG step 8 |
| bench-step8-triplet-{1,2,3}.md | Repeatability triplet (F15) | step 8 end-state |
| bench-step8-3c-*.md (4 files) | Contamination experiment: quiet brackets + download/hash arms | step 8 3c entry |
| bench-step8-block2-.md (4 files) | Qwen3-4B / Qwen3-8B / Llama-3.1-8B / DeepSeek-V2-Lite (F12) | step 8 Block 2 |
| bench-step8-block3a-*.md (2 files) | ubatch sweeps (pp512 + amended pp2048) | step 8 3a entries |
| bench-step8-block3d-*.md (2 files) | FA-on f16 and q8_0 KV depth curves (F16) | step 8 3d entry |
| bench-step8-thermal.log | Per-leg amdgpu edge temps, whole window | step 8 methodology |
| bench-step9-ws*.json (10 files) | Workhorse concurrency sweep, slots 1-16 x {f16,q8} (F17) | Step 9 results |
| bench-step9-oss20b-ws*.json (3 files) | gpt-oss-20b concurrency (knee at 4) | Step 9 results |
| bench-step9-coder30b.md | Qwen3-Coder-30B canonical bench | Step 9 gap-filler |
results/raw/ - survey + experiment raw outputs (76 files)
NOTE: step8/step9 files appear in BOTH raw dirs (identical copies - an artefact of two preservation passes). The build script deduplicates; checksums match.
| Pattern | Contents | Discussed in |
|---|---|---|
| bench-survey-*.md (12 files) | Tier 1+2 survey benches (phi4, mistral24b, gemma27b, qwen3-32b, r1-32b, phi4mini, coder32b, mixtral + mixtral-v2, scout-d{0,4096,8192}) | model-survey.md |
| bench-ladder-{q5km,q6k,q80}.md | Mistral-24B quant ladder rungs | model-survey.md Part 4 |
| bench-e1-scout-d*.md (6 files) | Scout chunk-boundary sawtooth sweep (-r 3) | experiments.md E1 |
| bench-e2-{q4,mxfp4}-ws*.json (6 files) | Flash quant-pair concurrency (F18 anti-scaling) | experiments.md E2 |
| bench-e4-scout-ws*.json (3 files) | Scout concurrency (best scaler) | experiments.md E4 |
| bench-e6-mixed-llm.json | LLM leg of the mixed-workload test | experiments.md E6 |
| bench-e12-ws{1,16}.json | Energy-economics serving windows | experiments.md E12 |
| bench-e13-*.json (5 files) | Spec-decode: 32B control, inert-draft, enabled-draft, workhorse draft legs (F21) | experiments.md E13 |
| bench-e14-*.json (3 files) | Co-residency: 4B solo + concurrent pair | experiments.md E14 |
| (step8/step9 duplicates) | see phase-a-raw table above |
results/eval-pilot/ - promptfoo outputs (23 JSONs + serverlogs)
| Pattern | Contents | Discussed in |
|---|---|---|
| {model}.json (5) | Pilot suite (17 tests) per model | eval-pilot.md |
| hardlong-{model}.json (5) | Hard long-context suite (6 multi-hop tests) | experiments.md E7 |
| e8-{model}.json (3) | 40.7K-token cross-chunk suite | experiments.md E8 |
| e9-{model}.json (5) | Contradiction-detection suite | experiments.md E9 |
| e10-{model}.json (5) | Structured-generation suite | experiments.md E10 |
| *.serverlog | llama-server -v logs per eval run (offload evidence, errors incl. the E8 context-overflow 400s) | E8 entry |
Test definitions and data: spikes/eval-pilot/ (promptfooconfig.yaml, hardlong.yaml, e8-32k.yaml, e9-contradiction.yaml, e10-generation.yaml, data/ - the invoice, log excerpts, 32K corpus).
results/flux-gallery/ - 50 generated images (gitignored, 67MB)
1024x1024 PNGs, seeds 1001-1050, prompts in flux-gallery-prompts.txt. Not in git; included in the package zip by the build script if present on disk.
Publication status
NOT YET PUBLISHED. Pre-publish checklist (QUESTIONS-FOR-ALASTAIR Q4): license for data + images, venue (no-github rule applies), content scrub sign-off (eval data embeds internal log excerpts - they ARE the published data, but confirm deliberately).