Community recipes and hardware submissions

The version-1 registry contains five maintained recipe records: policy- conditioned alignment, a preference-teacher distillation template, localized verifier-gated OPD, tool-policy alignment and RecoveryBench. Every entry records the exact model revision, GPU/VRAM, wall time, benchmark, artifact hash, measurement status, compatible miniVERL release and bridge profile. Four initial records bind maintainer measurements already preserved in this repository. The preference-teacher entry is explicitly not_measured: its linked aligned-adapter recipe is a scaffold that still names the public SFT teacher, and users must substitute a pinned preference-trained adapter before claiming that category. None is presented as external adoption.

Create a privacy-safe unmeasured submission template without installing torch:

miniverl benchmark --export-community submission.json

The schema requires hardware/software provenance and a packaged recipe digest. Measured submissions additionally require wall time and artifact hashes. The validator rejects recipe mismatches, credentials, environment references and local absolute paths. Change measured_status only after retaining the source run and matching its published hashes.

To propose a real result, add the JSON plus any new versioned recipe record in one pull request. Describe unsupported hardware or software explicitly; a negative or failed run is useful evidence and should not be removed merely because it does not improve a benchmark.