Evidence index: 2026-09-26 single-GPU qualification¶
This closed bundle records a no-promotion decision. Native suite artifacts are
retained under raw-run-evidence/; their schemas and result counts are not
flattened. The two sanitized SWE files retain their native schema but replace
operator paths and transient container identifiers. Recipe reconstructions use
${MODEL_CACHE} and ${ENGINE_CACHE} placeholders.
- Decision summary, source registry, workload and gate contract, and coverage.
- Restoration, friction log, and redaction ledger.
- Closed artifact ledger. The publication-summary role is not applicable because no compact public communication was requested.
ThinkingCap coding raw evidence is missing after managed cancellation. The retained review excerpts preserve no complete trajectory, patch, grader output, or completed score.
Native gate reports¶
| Configuration | Agentic / context | Images | Coding |
|---|---|---|---|
| Flash IQ3 r3 | 18/18 / 12/12 | 12 image, 4 eight-image | 3/5, all graded |
| Swift27 r2 | 18/18 / 12/12 | 12 image, 4 eight-image | 2/5 attempted, 4 graded |
| ThinkingCap r1 | 18/18 / 12/12 | 12 image, 4 eight-image | incomplete; observed excerpts |
The Flash preflight includes the 20-request tool gate. GLM reuse verification binds the retained 4/5 reference.
Capacity diagnostics retain C1 unique, C4 shared, and the invalid C4 unique timing with its cache-history correction. The correction takes precedence over the native artifact's eligibility flag. These populations do not support a qualified model-speed ranking.
Restoration checks are retained as direct and authenticated routed native reports, alongside the exact identity in the restoration receipt.