GLM v0.9.1 trial evidence

This public-safe bundle records the completed R2 and restored-incumbent evidence for the 2026-10-16 GLM v0.9.1 trial. Both completed the matched five-case SWE control at 3/5, below the four-case replacement gate. The candidate remains rejected with no promotion; the incumbent is restored. The publication record has passed independent review.

• Candidate R2 summary: functional and quality gate results, including the failed 3/5 coding gate.
• Restoration summary: exact incumbent restoration and post-run acceptance evidence.
• Sanitization provenance: private-source hashes, derived public files, and redaction boundary.
• Artifact manifest: closed inventory and artifact roles.
• Source registry: upstream recipe inputs, kept separate from local qualification evidence.
• Native artifacts and derived indexes: source-schema-preserving SWE artifacts plus explicitly labeled derived indexes for repeated counts.
• Coverage and gaps: gate status and explicit control dependency.
• Independent baseline review: the allowed admitted-control comparison boundary.

The candidate has no eligible performance result. It was rejected for the declared replacement gate, with no promotion or route change. The incumbent's fresh bounded control does not make the R2 coding failure pass.