cbuk10001/Cam1-decision-index

Cam1 — complete Decision Index 0.3 Public evaluation

The submitted evaluated engine scores 30.72 Public points, with official scores.json reporting complete: true. The complete suite has 140,620 executions, 140,178 scoreable requests, 43 benchmarks and 37 Public-index contributors. Of the scoreable requests, 138,071 succeeded and 2,107 were unsupported; there were zero runtime errors. The official 442 exclusions are unchanged.

Model and runnable inference code: cbuk10001/Cam1. The results belong to the evaluated cam1_engine:Cam1Engine behavior, originally run as official_engine:OfficialNativeEngine from source commit 32232f087ed3ae72f278ed78f2e9dd325eed2835. The suite, runner and scorers are unchanged from official commit 9eb2dbe2a358004c8782c66e40a83ac07b953fec.

Area	Public area points
Knowledge & Reasoning	19.21
Language Understanding	33.84
Retrieval & Classification	36.92
Tools & Automation	46.33
Arts & Human Taste	11.45

These are chance-corrected, coverage-adjusted index points, not raw accuracy. This Public result is not the maintainer-run Full score, an official rank, or admission certification. Full and held-out latency evaluation remain the maintainers' responsibility.

Results integrity

runs/Cam1-evaluated-fp32-v1/results.jsonl is the official compact storage projection of the completed original results. Only payload and raw_output were removed, as documented for the runner's --compact representation. Every retained field, answer, status, request ID, timestamp/timing and row order is identical. Original raw results remain independently archived privately. This storage-only projection avoids redistributing licensed evaluation text; it does not substitute predictions or alter scoring.

The original scores.json, benchmark-summary.json, index.json, environment.json, run-manifest.json and integrity.json are copied without modification. Independent local rescoring of the compact export exactly reproduced all scores, benchmark summaries and index values; only the scorer's newly generated timestamp differed. See compact-export.json and rescore-verification.json for identities and verification. The score file keeps the original historical engine/run labels; the public directory is named Cam1.

Capacity and timing

One NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB VRAM; Python 3.12.15; Torch 2.11.0/CUDA 13.0; Transformers 5.10.2; PEFT 0.18.1. Full FP32, math SDPA/reference kernels, deterministic execution, TF32/autocast disabled, GPU-only, one process/request, maximum 16,384 tokens. The carried temperature is 1.333521432163324; it was not fitted on benchmark outcomes. Its original BF16 calibration is not proven valid under FP32.

The scoreable refusals are 107 overlength requests (25 ToolRet, 81 BRIGHT, one POP909) and 2,000 PhishNChips schema refusals (empty Noul criteria/null Choice descriptions). An additional 396 overlength executions are in the official exclusions. Nothing was truncated and no options were removed.

Across the 138,071 successful scoreable requests, synchronized model-request wall latency was mean 408.97 ms, median 255.28 ms, p80 398.75 ms, p95 915.28 ms. Scoreable throughput including refusals was 2.482 requests per summed request-time second. Peak allocated/reserved GPU memory across attempts was 54.79/86.25 GiB. These are a preview; maintainers independently require held-out mean, median and p80 each at most 1,000 ms.

The run was interrupted and resumed with identical runtime/artifact identities. environment.json records the resumed process; its peak-memory fields and attempts/resume/ telemetry/status/completion are resume-only, not whole-run totals. combined-performance.json contains both attempts and the whole-run aggregates. The complete evaluation cost is estimated at $43.55; it is not a provider invoice and excludes later diagnostic work. Per-benchmark scores are in benchmarks.csv and the official JSON files.

Separate post-evaluation Phish compatibility diagnostic

The model repository also exposes Cam1CompatibilityEngine, accepting empty Noul criteria and null Choice descriptions generically. Weights, tokenizer, prompt construction, temperature and numerical policy remain unchanged. The fix answered all 2,000 Phish inputs, scoring 1,024/2,000 (51.2%) verdict accuracy. All 507 regression requests were identical, including raw logits/responses on 504 supported requests and three identical overlength refusals; all 32 fixed FP32 fixtures replayed exactly.

This diagnostic is separate from the submitted complete run. No corrected Phish results were spliced into the full results, and no new complete corrected score is claimed. See compatibility-diagnostic.json. The default evaluated entrypoint retains the original refusals so maintainers can reproduce the submitted behavior.