CLAIMBOUND / SOFTWARE PERFORMANCE
Narrow, pre-registered Apple Silicon comparison. ManifoldBT Community CPU only versus open-source Monte-Neo (MIT). CUDA / Pro GPU is not a peer on this host.
PASSED_UNDER_PROTOCOL under MANIFOLD_MONTE_NEO_FAIR_RACE_D001 (registry sequence 8073).
Operator: NeoZorK · Host baseline: Apple M1 Pro 16GB · Synthetic OHLCV seed 42, 50k bars.
| Lane | ManifoldBT | Monte-Neo | Gate |
|---|---|---|---|
| Type C — sweep (combos/s) | ~38 | ~46k | Monte-Neo > Manifold (required) |
| Type A — return-path MC (sims/s) | ~3.2k | ~1.9k | both ok (no winner) |
| Type B — scenario / N-run | ok | Metal or CPU fallback | informational |
Type C label: specialized Numba SMA kernel (Monte-Neo) versus
general-purpose run_sweep (ManifoldBT). This is the D001 throughput gate,
not a matched fee-aware full-engine bake-off.
Type A/B/C answer different questions and must not be merged into one speed claim.
Relative paths into the evidence tree (also on GitHub):
runner_command.claimbound-evidence, run that command.PASSED_UNDER_PROTOCOL and Type C gate still passes (absolute rates may vary by host).
Card:
CLAIMBOUND-MANIFOLD_MONTE_NEO_FAIR_RACE_D001-2026-09-12.json
Protocol:
MANIFOLD_MONTE_NEO_FAIR_RACE_D001 charter
Independent rerun issue:
#355
Does not claim universal superiority of any engine. Does not compare against Manifold CUDA/Pro GPU. Does not claim Type C is an apples-to-apples full-engine comparison. Community caps apply (sweep ≤256, MC ≤1000). Synthetic data only.