GLM-5.3 NVIDIA FP4 controlled A/B

The initial candidate context result was 17/18, versus incumbent 18/18. After the user authorized investigation, the unchanged candidate scored 18/18 on a full replay and passed all three follow-ups of the originally failed case. The failure did not recur; its cause remains unproven. No configuration repair, performance win, qualification or promotion is claimed.

The exact original r11 container is restored. Direct and authenticated routed smoke/JSON checks passed2/2 each after the final restoration; original configuration hashes are unchanged and the candidate service is absent.

• Decision and limits, diagnostic result, restoration.
• Configuration, workloads, initial run plan, diagnostic plan.
• Native evidence: baseline, initial candidate, first restoration, diagnostic and final restoration.
• Verification, artifact inventory.
• Sources, cause investigation, fixture reconstruction, failures, coverage.

Initial failed artifacts remain retained. Follow-up observations are separate; matching reconstructed fixtures does not prove the original transmitted bytes, and the small sample with different request/cache histories does not establish a general failure rate. Candidate timing, overlap/endurance and repository-scale coding remain untested. The conditional MTP-off arm was not run because the failure did not recur.