7132997511
* feat(bench): add Apple Silicon UMA mechanical sympathy benchmark Engineering-grade reproducible benchmark measuring exact CGA recall, Cl(4,1) scalar algebra, FrameVerdict TTFV, array_codec replay, and honest Python/Rust copy/zero-copy boundaries. Runs without Rust; skips Rust-only tracks with explicit reasons. Includes claim-safety audit, CLI integration (core bench --suite apple-uma), and outreach brief. * fix(bench): patch apple-uma report paths, decode timing, CLI --report - Use repo-relative report_path in JSON metadata (no absolute paths) - Measure decode_array only; precompute encode payload before decode bench - core bench apple-uma --report writes exactly to PATH; --write-report for defaults - Add final newlines; regenerate seed report
2.7 KiB
Apple Silicon Engineering Support Brief (Draft)
This brief is factual and engineering-first. It does not claim Apple endorsement,
review, or sponsorship. All performance and memory-boundary claims are backed by
the reproducible benchmark report at
evals/reports/apple_uma_mechanical_sympathy_latest.json.
What CORE is
CORE is a deterministic Cl(4,1) reasoning and safety engine. Its runtime path preserves geometric field invariants, exact CGA recall, closed-world proof surfaces, and replay-stable evidence — not stochastic token generation.
What the benchmark measures
The CORE Apple Silicon UMA Mechanical Sympathy Benchmark measures:
- Exact CGA top-k recall throughput on contiguous
(N, 32)float32 matrices - Cl(4,1) scalar algebra hot paths (
geometric_product,versor_apply,cga_inner,versor_condition) - Off-serving closed-world FrameVerdict time-to-first-verifiable-verdict (TTFV)
- Deterministic
array_codecpersistence replay cost - Honest Python/Rust copy and zero-copy input boundaries
It does not benchmark token generation, approximate recall, or transformer throughput.
Why Apple Silicon UMA is relevant
CORE workloads are dominated by contiguous-memory geometric operations and exact
recall scans. On Apple Silicon unified memory architecture, native bindings that
avoid Python marshalling tax on hot paths (for example Rust vault_recall and
diffusion_step input views) align with mechanical sympathy for UMA — when
measured, not assumed.
Current hardware limits
On the machine that generated the latest report, larger validation lanes (for
example N=65536 exact recall, large diffusion graphs, expanded replay buffers)
may be skipped or constrained by available memory and single-node throughput.
The benchmark records these limits explicitly.
Requested engineering feedback
We are seeking Apple Silicon engineering feedback on:
- Whether measured UMA-aligned workloads match expected memory behavior on M-series
- Practical guidance for MLX/Metal kernel experiments under separate ADR/parity gates
- Whether expanded hardware access would unlock larger reproducible validation runs
Future-facing context (not benchmark claims)
Deterministic verification and replay throughput may be relevant to on-device safety and audit surfaces in future R&D — but that relevance is not claimed as a current product integration. MLX, Metal, CoreML, and Neural Engine paths remain future work until implemented, parity-tested, and measured.
How to reproduce
python -m benchmarks.apple_uma_mechanical_sympathy --write-report
# or
core bench --suite apple-uma --write-report
Reports land under evals/reports/. No network access is required.