METALBENCH

METHOD / V1

Receipts before rankings.

The rules are deliberately boring: deterministic grading, visible configurations, and no winner when the inputs don’t match.

Grading

Benchmark tasks use exact matching, substring checks, JSON parsing, or another declared deterministic rule. An LLM does not judge the answer.

Provenance levels

Documented local
Transcribed from dated local notes. Useful, but not independently reproducible without the raw artifact.
Artifact-backed local
Includes a sanitized JSON artifact, SHA-256 digest, and harness commit.
Published external
Links to an HTTPS publisher source. Never determines a local winner recommendation.

Comparison policy

Hardware, benchmark version, task-set hash, and concurrency must match. Runtime, quant, context, and reasoning differences remain visible warnings.

Memory-fit rule

Total weights must fit within available memory with room for runtime overhead, KV cache, the operating system, and co-resident services. “The model file fits” is not the same as “the workload fits safely.”

Privacy claim

Artifacts and the final build pass an automated scanner plus manual review. This lowers risk; it is not an absolute guarantee that no sensitive detail can ever escape.