[BANADDA]

One epoch now runs as the operator's validator flow, one function per step.

Setup, automatic and once per arena: load the epoch; reference parts and allowed bases from the cache or downloaded when missing; the probe agent placed from the GPUs the validator sees (its own GPU when there are two or more; sharing one card of 40 GB or more with the full model; sharing a smaller card on testnet with the AWQ build of the same model; refused otherwise); the reference, frontier redo and each base run on the hidden items, with a failing frontier redo marking those pools as having no reference instead of stopping the epoch; the miner queue built from metadata only, ordered by commit block.

Per miner loop, one at a time or one per GPU: fetch and check, evaluate (SLM, harness, escalation, full system per floor), score, then post the signed miner result to POST /v2/reports/artifact and delete the artifact and SLM weights. A failing miner is posted with its reason and deleted the same way. Only router folders are kept, moved aside.

Close: the router pass scores each kept router on the other miners' and the reference drafts, posts again with the router pool and deletes the folder; failed posts are retried; weights; the signed epoch report; probe sets; the work folder cleared except the report; base weights stay cached.

The sandbox is a locked down Docker container everywhere (no gVisor default, no entrypoint check), missing sibling images are pulled with retries, the manual judge settings from PR 170 are replaced by the automatic placement, and a malformed manifest is a per miner refusal instead of an epoch failure. Docs updated.

🤖 Generated with Claude Code