NVIDIA FP4 promotion evidence index

This is the sanitized documentation view of the human-approved primary selection on 2026-10-17. The operator considers this pinned recipe the best current fit for this deployment. Full native promotion/client captures remain private; their summary hashes are retained in summary.json.

Record	Purpose and boundary
Promotion finding	Selection rationale, exact configuration, acceptance, KV interpretation, rollback, and incomplete fleet gates
Promotion summary	Machine-readable current decision, acceptance/source hashes, client and publication status
Initial promotion summary	Immutable earlier fleet-state snapshot, superseded by supported recovery in the current summary
KV accounting	Configured byte pool, measured C4 block fractions, estimated occupied/unused GiB, and source hashes
Full serving comparison	Authoritative comparison of the FP4 and EXL3 recipes across 12 matched pairs
Comparison tables	p50/p95 timing, prefill, decode, TPOT, throughput, tokens, memory, KV, and concurrency
Comparison evidence index	Original plan, native results, telemetry, failures, and historical restoration
Comparison artifact manifest	Hash-bound native evidence inventory; the original comparison's no-promotion close is historical

The fresh promotion capacity population has 12 scored requests after one excluded warmup; the earlier natural/edge comparison has 84 scored requests per arm. Native client smokes are functional evidence and do not remeasure engine timing or KV. Their unavailable metrics and excluded captures remain explicit. No new inference ran to prepare this documentation.

The older workflow is now cancelled and verified quiescent. Corrected native profiles passed staging, while full catalog/session acceptance and the declared secondary dependency remain pending. The promoted primary and retained rollback are separate verified states; documentation success does not prove fleet acceptance or four simultaneous full-window decodes.