GLM serving comparison evidence — 2026-10-17

The finding explains the complete matched natural-output/capacity-edge scope and the failed controlled-output scope. This is a local documentation artifact set, not a model promotion or fleet release.

• All timing and resource tables
• Exact numeric comparison
• Seven-metric chart
• Decision and exact coverage
• Native comparison envelope
• Frozen comparison plan and request run plan
• Independent measurement review
• Paired controlled failure
• Original timeout adjudication
• Collection recovery
• Final restoration and independent review
• Source/redaction provenance
• Closed artifact inventory

All 24 accepted cells retain the 17 required metrics plus four supplemental families, native responses, per-request timings and in-flight telemetry. They contain 12 matched pairs and 84 scored requests per arm. Six warmups, two failed one-request diagnostics, the original controlled timeout and excluded attempts are retained separately. The frozen 42-cell plan remains closed partial: 24 accepted, one failed, 17 blocked, zero pending. No missing value is imputed.

The public bundle sanitizes operator identities, paths, endpoint ports and GPU identifiers, while retaining numerical measurements, request populations and model/image revisions. Structured references bind the sanitized public bytes; private-origin-provenance.json preserves the source-to-public hash map. Compressed telemetry is deterministic gzip. The original timeout's large stream uses ordered whole-record chunks and a concatenated sanitized-stream hash.

Eleven historical JSON snapshots are explicitly archived as .json.txt with original-source hashes. They preserve superseded evidence and are not current required-reference or completeness proof. Their old bindings may be stale. The current evidence graph is separately validated. Frozen source and collector snapshots document the actual measurement producer; later source safeguards were applied only after final restoration. Sanitized source snapshots are documentary evidence, not executable installations: redaction can alter syntax. The frozen source revision and original file hashes identify the executing code; the corresponding public product source is available at that Git revision.

The authoritative model repositories and exact revisions are pinned in the comparison plan: NVIDIA FP4 and EXL3. External recipe guidance is not local performance proof. The earlier quality campaign remains a separate artifact set.