Prism

Prism is a new competition: prism-v2.1, recipe 2.1.0, scoring generation 21. Only harvests finalized under that contract can receive leaves, rank, or take weights.

Recipe 2.0.0 and 1.x scores are dead. They are not rescored, cannot win WTA, and cannot be paid. Re-submit a 2.1 AutoModel patch if you want to compete.

Until the first eligible 2.1 submission terminates, subnet weights stay burn — uid 0 = 100%. An empty board is honest, not a missing API.

GET /v1/weights/latest is fail-closed: with no sealed 2.1 bundle the gateway serves the burn vector (sealed: false) rather than inventing a champion.

Submit a ZIP (preferred) — or JSON with the same members / zip_base64. Recipe 2.1.0 accepts only an AutoModel pin id plus your git diff against that pin. This is not a free-form architecture.py / training.py project.

Workflow: read GET /v1/recipe (live pin automodel@v0.5.0), check out that commit, edit, git diff the pin, write automodel.base as that pin id, pack the ZIP, POST /v1/submissions with your hotkey and X-Lium-Api-Key.

automodel.base          # required — pin id from GET /v1/recipe (live: automodel@v0.5.0)
automodel.patch         # required — unified diff vs that pin
prism.toml              # optional — entry / model-config knobs
requirements.txt        # optional — one of requirements.txt or pyproject.toml

Live eval is miner-funded. Create a Lium account, fund it, and send X-Lium-Api-Key on every live submit. Missing key on live → 400 missing_lium_api_key. The key is held in master memory for that submission and may land in a TTL-bounded encrypted seal (default ≥36h); never in Postgres, never logged.

Live eval pins 1× NVIDIA B200. If Lium has no matching offer, intake still accepts the ZIP (202 + a note) and the run stays queued. B200 sold out ≠ rejected and is not Score(0). The orchestrator retries rent on the next worker tick. GET /v1/status advertises the same policy as lium_capacity_note. Auth, a missing X-Lium-Api-Key, a bad ZIP, or a template-permission failure after Lium’s own fallback still fail — those do not queue.

You may ship one of requirements.txt or pyproject.toml at the AutoModel repo root in your patch. It installs in a network-on phase before your model code is sandboxed (train/eval stay offline).

Models must stay in 850M–1B parameters (total unique; tied embeddings once). A 215M pack is invalid for recipe 2.1. Trust GET /v1/recipe for min_params / max_params, train_hours_cap 4.0, and the live pin — not marketing chart labels.

Legacy 1.x two-script ZIPs and closed 2.0 harvests cannot win. Do not ship Megatron-Bridge or other non-AutoModel frameworks. HTTP only — no Phala/CVM.

Live scoring_version 4 ranks by the equal-weight mean of available G2 public-suite accuracies (HellaSwag, ARC-E/C, PIQA, WinoGrande, BoolQ, …) mapped to round(SCORE_MAX × mean). Bits/token bpb is recorded for display / G1 and does not drive emission.

Cheat / suspicious agentic gates hard-zero the leaf. Emission is winner-take-all on your own best 2.1 training score. Until someone finishes an eligible 2.1 run, weights burn.

Normative miner guide: BaseIntelligence/prism. Control-plane contract: docs/external-miner/prism.md.

# Inspect the live 2.1 pin
curl -sS https://chain.joinbase.ai/challenge/prism/v1/recipe

# ZIP via the production gateway
curl -sS -X POST https://chain.joinbase.ai/challenge/prism/v1/submissions \
  -H 'content-type: application/zip' \
  -H "X-Miner-Hotkey: $MINER_HOTKEY" \
  -H "X-Lium-Api-Key: $LIUM_API_KEY" \
  --data-binary @submission.zip