Publication validation
- 128 experiment-package tests passed, including bounded delivery, same-adapter final requests, strict evidence commitment, semantic review, shared-draft provenance, task cleanup, export normalization, adapter identity and ordinary-runtime opt-in behavior. The official SciBench scorer comparison ran with the local upstream evaluation tool available.
- 51 related MiroFlow tests passed across code-interleave policy, LoRA runtime MCP, prompt assets and SGLang routing. One inherited prompt assertion was aligned with the existing baseline wording that discourages repetitive plan/search cycling; the ordinary routing prompt itself was not changed for that assertion.
- 420 exported trajectories / 14 groups verified, with zero recorded-score mismatches and 180 identical paired first requests across the six passed workflows. Counts, terminal submissions, indices and archive hashes were checked. The legacy overall workflow remains incomplete.
- The portable preparer was exercised against the recorded adapter hashes. Frozen source hashes, Hydra profile loading, MCP configuration resolution and refusal to overwrite an existing run root passed. The launcher passed Bash syntax checking. Ordinary MiroFlow imports were also checked without the experiment package on
PYTHONPATH. - Publication files were inspected for credentials and local deployment identifiers. Metadata normalization preserves opaque identities and mathematical program variables.
checksums.jsonin the evidence bundle identifies the published files.
No new GPU benchmark was run to assign a score to this publication refactor. Historical native startup/A/A/workflow evidence is supplied separately, and a fresh deployment must pass its own native gates. The local test environment emitted an optional audio dependency warning about ffmpeg; the tested scientific-code paths did not use audio.