Skip to content

GLM-5.3-Flash APC campaign evidence

This public bundle retains sanitized local evidence for the 2026-09-19 APC comparison. It uses generic direct endpoints: baseline http://127.0.0.1:30000/v1 and candidate http://127.0.0.1:30001/v1. Operator paths, host identities, GPU UUIDs, container IDs, and command traces are excluded.

Both 32-request finalist pairs are complete. At nominal 32K/C4 (about 25K observed prompt tokens), APC's unique-cache request throughput was 2.388% lower, within the 5% non-regression gate; shared-cache visible TTFT fell 67.84% and request throughput rose 3.1655x. The earlier eight-request scout remains a negative descriptive signal. The candidate also passed four eight-image cases, a 309,422-token needle, a 110,760-token tool call, and bounded diagnostic quality checks. These are not a SWE/coding-intelligence result. Baseline restoration and direct/authenticated routed acceptance are verified. No route or model promotion is authorized by this evidence.

Final comparison, reasoning-control provenance, independent review, and failed original 64-word cell are retained.