Skip to content

SGLang, Qwen, and Nemotron comparison evidence

This public-safe bundle is the evidence index for the 2026-10-16 bounded scout campaign. It preserves native schemas and terminal failures without publishing raw prompts, model output, trajectories, tool arguments, commands, endpoints, private paths, host identity, GPU UUIDs, or container identities.

Decision

no-promotion. The original GLM baseline and Nemotron both have C1 0/4 performance-eligible requests; SGLang GLM and Qwen FP4 miss the frozen coding floor, while Nemotron also misses the campaign exact Oracle gate. No throughput ranking or graph exists.

Retained evidence

  • baseline-public-configuration.json, sglang-scout-public-configuration.json, qwen-scout-public-configuration.json, and nemotron-scout-public-configuration.json bind public-safe identities and scope.
  • native/ preserves preflight, Oracle, context, vision, capacity, SWE, and failure projections in their native schemas.
  • source-registry.json, run-plan.json, and workload-manifest.json preserve inputs and the late common-packet initialization.
  • summary.json, friction-log.md, and restoration.json are campaign-level decision, process, and restoration records, including the pre-trial router-custody checkpoint and its bounded caveats.

The dated finding is the authoritative narrative. Artifact manifest is the closed inventory generated from artifact-manifest-source.json after final restoration.