The question

Can a private intelligence system avoid redundant work while preserving model, identity, permission, and artifact checks? Cavumo compiles a profile into a bounded plan and reuses only state whose validity can still be established.

The experiment

We compared Cavumo’s fixed configuration with guarded compiled execution using the same four-bit Gemma 3 1B IT model, M2 Mac with 8 GiB memory, binary, decoding settings, and context budgets. A frozen 340-query synthetic workload covered documents, memory, and shared context. Three counterbalanced paired repetitions produced 2,040 generations and 1,020 pairs. Timing included Service context resolution, planning, admission, integrity checks, fresh worker execution, and streamed Metal output; activation preflight was excluded equally.

What we observed

Overall p95 latency fell from 4,983.29 ms to 2,669.39 ms: 46.43% lower, with a paired bootstrap interval of 32.36%–62.51%. Generated outputs matched in every pair. Cold-runtime p95 improved 7.41%, warm 65.92%, memory 12.44%, and shared context 9.85%; cold and shared confidence intervals include zero. This is a scoped systems improvement against our own fixed configuration, not a comparison with cloud APIs or other runtimes.

Quality and limits

Exact synthetic answer correctness was 64.12% in both conditions, and abstention on unanswerable cases was 96.10%. Generated citation support was not measured. These fixtures predominantly require short factual answers; they do not establish customer-domain quality. Runs used a headless development profile on an interactive desktop, not an isolated performance lab. Repeated-query confidence intervals do not describe between-customer uncertainty.

What comes next

The study passed its frozen overall compiler target. Later correctness changes and the eventual production build require separate performance validation. The compiler is an execution planner, not a new GPU kernel, CUDA replacement, or evidence of patent novelty.

Evidence & reading copy

Download the scoped study or engineering note. Figures are descriptive; they do not run the product in your browser.

compiler-study.json Full note · Markdown