Skip to content

ADR-0008 raw-evidence gap — 2026-07-22

  • Related: ADR-0008, ADR-0027, issue #175
  • Evidence type: public provenance correction / missing-artifact record
  • Historical decision date: 2026-07-02
  • Correction date: 2026-07-22

Gap

ADR-0008 reports a live NEXTN speculative-decoding A/B on Qwen3.5-35B-A3B AWQ, including throughput, TTFT, draft acceptance, cache-hit, error, and restart observations. Its raw benchmark logs were not committed to the public product repository.

The #175 audit fetched fakoli/anvil-serving-notes through 7b46ceb6ae62252f8f808f6c065706a24e7970bb and searched all refs for the ADR's exact throughput values (187.81, 268.15, 486.16, 630.64), 96.2% cache-hit observation, Qwen3_5ForCausalLMMTP, and the four-draft-token flag. No matching narrative or raw artifact was present. No prior byte size or SHA-256 digest is available because no artifact path was recorded.

Consequence

The measurements remain an attributed historical observation in the ADR, but external readers cannot independently recompute them. They must not be used as current qualification or promotion evidence for another checkpoint, engine build, quantization, GPU topology, or routing policy. Equivalent future decisions require a fresh sanitized narrative and bounded raw artifacts under ADR-0027. This record does not fabricate replacement evidence or imply private access can fill the gap.

No product or serve configuration changes as a result of this provenance correction.