September 23 model qualification evidence

The native artifacts preserve successful, failed and partial attempts. This bundle supports a bounded no-promotion decision for GLM Spark and Qwen Flash Next on two RTX PRO 6000 Blackwell Max-Q GPUs. It is not a full SWE-bench score, a general model ranking or a long-session reliability guarantee.

Campaign and decision

• Machine-readable summary: results separated by exact request profile.
• Artifact manifest: closed file inventory, ten evidence roles and hashes.
• Source registry: dated official and community leads; archive screening is not independent confirmation of every report.
• Coverage and remaining limits.
• Failures and durable dispositions.
• Restoration receipt.

The user authorized conditional promotion after qualification. Neither challenger passed the required gates; no alias was changed. Qwen default xhigh and medium are separate profiles. Native oracle thresholds are preserved in raw files; the campaign requires all 18 deterministic assertions.

Reconstruction and workload

• Incumbent recipe, GLM Spark, Qwen r1, Qwen r2.
• Frozen campaign plan and workload manifest.
• Five-task selection: one task per repository selected before outcomes, 60 calls and 16384 completion tokens per call.
• Long-guide suite: one natural generation, 65536 total completion-token budget, including reasoning.
• Verified oracle-7 launcher. Earlier launcher and oracle records remain historical evidence.

Sanitized recipe paths and device identities are examples to replace through managed recipe controls. No recipe is permission to evict an active workload. Configured maximum context/concurrency is distinct from measured capacity.

Native evidence groups

native/ retains native preflight, agentic, multimodal, capacity, quality and SWE schemas. Each profile in the summary names its exact source files. Official SWE reports are retained for submitted patches. Unsubmitted attempts remain visible and count in the all-attempts denominator. Per-instance token and duration values absent from native SWE evidence remain unknown.

review/ retains independent critiques, static long-guide screens, startup limitations, test/source receipts and measurement confounds. These supplement rather than replace the native results. Marker and syntax checks do not prove that generated chapters form an integrated working program.

Strict capacity cells use unique prompts and request canaries. Any count, marker, truncation or completion failure makes that request ineligible for performance claims. No throughput chart is produced from failed populations. Small samples do not establish meaningful p99 latency or causal MTP speedup.

Redaction and provenance

Sanitization receipts bind original private bytes to public copies. Real operator paths, hosts, container IDs and GPU UUIDs are replaced with synthetic examples. Nested source hashes in native artifacts still identify the original source bytes; the campaign manifest hashes the public copies. No private credentials or user task/session content is published. Retained model outputs belong to the synthetic benchmark workload.

Public social copy was not requested, so the publication-summary role is not applicable. The dated finding and documentation matrix provide the publication.