AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
翻譯準備中
SemanticCompute — public distribution: signed binaries, docs & licence for the numerical-verification tool. Source is closed (commercial licence).
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Numerical verification for heterogeneous compute. Bring any two arrays — a CUDA, Metal, Triton, or
hand-written kernel result and the reference you trust — and prove they agree under a tolerance you state, or get
an exact account of where and why they diverge. SemanticCompute classifies FMA contraction, denormal flushing,
NaN/infinity differences, reduction order, and ULP drift, and measures accuracy against a higher-precision truth
where a family supplies one. Verification is the product; SemanticCompute is not a general GPU framework.檢視完整 README ↗目前無法載入完整 README。請透過原文連結閱讀。 以下為已儲存的摘錄。
Numerical verification for heterogeneous compute. Bring any two arrays — a CUDA, Metal, Triton, or
hand-written kernel result and the reference you trust — and prove they agree under a tolerance you state, or get
an exact account of where and why they diverge. SemanticCompute classifies FMA contraction, denormal flushing,
NaN/infinity differences, reduction order, and ULP drift, and measures accuracy against a higher-precision truth
where a family supplies one. Verification is the product; SemanticCompute is not a general GPU framework.檢視完整 README ↗