BASE Documentation

This guide is for PRISM challenge operators (running the PRISM service). It is not the BASE subnet validator pack under Validators. This guide covers local validation and production-oriented configuration for running PRISM as a BASE challenge. Source: docs/operators.md:1-5.

Installation

git clone https://github.com/BaseIntelligence/prism.git
cd prism
uv sync --frozen --extra dev

Source: docs/operators.md:6-12 (repo URL per SOURCES.md:51).

Local validation

.venv/bin/ruff check src
.venv/bin/mypy src/prism_challenge/evaluator
.venv/bin/python -m pytest tests -q

Source: docs/operators.md:14-20.

Required runtime configuration

At minimum, PRISM needs a database URL, a shared-token file, and an execution backend:

PRISM_DATABASE_URL=sqlite+aiosqlite:////data/prism.sqlite3
PRISM_SHARED_TOKEN_FILE=/run/secrets/platform/challenge_token
PRISM_EXECUTION_BACKEND=platform_gpu

The shared token must match the token configured in the master for this challenge. Source: docs/operators.md:22-32.

Docker broker configuration

Production evaluation uses the Docker broker with the augmented evaluator image:

PRISM_DOCKER_ENABLED=true
PRISM_DOCKER_BACKEND=broker
PRISM_DOCKER_BROKER_URL=http://platform-docker-broker:8082
PRISM_DOCKER_BROKER_TOKEN_FILE=/run/secrets/platform/challenge_token
PRISM_PLATFORM_EVAL_IMAGE=<deployment-specific evaluator image>
PRISM_PLATFORM_EVAL_GPU_COUNT=1
PRISM_DOCKER_NETWORK=none

Set PRISM_PLATFORM_EVAL_IMAGE to the evaluator image published for your deployment. The image must ship sentencepiece and an offline tiktoken gpt2 cache so reference tokenizers load with no network.

The scored run is single-node and uses torchrun --standalone --nnodes=1 --nproc-per-node=1. Source: docs/operators.md:34-50 (evaluator image literal omitted; it references a non-BASE registry - supply your own).

Locked FineWeb-Edu data plane

The broker bind-mounts the locked FineWeb-Edu data read-only into the eval container, which runs with network=none:

PRISM_PLATFORM_EVAL_DATA_DIR=/data/fineweb-edu/train       # miner-visible, read-only
PRISM_PLATFORM_EVAL_VAL_DATA_DIR=/data/fineweb-edu/val     # secret; scorer-only, never mounted into eval
PRISM_PLATFORM_EVAL_REFERENCE_TOKENIZER_DIR=/opt/reference-tokenizers

HF_HUB_OFFLINE=1 and HF_DATASETS_OFFLINE=1 are set inside the eval container. The val/test splits are secret and must never be exposed to a miner script. Source: docs/operators.md:52-64.

Compute budget

The score is compute-normalized; wall-clock is only a safety cap, enforced in layers:

PRISM_PLATFORM_EVAL_BUDGET_SECONDS=1200            # graceful stop; score the partial stream
PRISM_PLATFORM_EVAL_WATCHDOG_GRACE_SECONDS=120     # hard watchdog above the graceful budget
PRISM_PLATFORM_EVAL_TIMEOUT_SECONDS=1800           # outer docker/broker backstop
PRISM_PLATFORM_EVAL_ARTIFACTS_QUOTA_BYTES=2147483648

Source: docs/operators.md:66-75.

LLM hard gate configuration

The OpenRouter LLM hard gate is enabled by default and reviews both scripts before any GPU work:

PRISM_LLM_REVIEW_ENABLED=true
PRISM_OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
PRISM_OPENROUTER_MODEL=openai/gpt-4o
PRISM_OPENROUTER_API_KEY_FILE=/run/secrets/openrouter_api_key

A reject from the gate is terminal. The eval container carries no OpenRouter key (the gate runs host-side before the container is launched). Source: docs/operators.md:77-89; config.example.yaml:60-63.

Multi-GPU static contract

PRISM_DISTRIBUTED_CONTRACT_POLICY=reject     # reject | flag | off
PRISM_PLATFORM_EVAL_MAX_GPU_COUNT=8

reject (the default) hard-rejects a non-distributed training.py; flag advances but logs; off skips the check. Source: docs/operators.md:91-99.

Duplicate review

PRISM_PLAGIARISM_ENABLED=true

An exact-source-hash duplicate is rejected, and a borderline-similarity quarantine is folded into a terminal rejection at ingress. There is no operator hold-resolution surface. Source: docs/operators.md:101-108.

Running locally

PRISM_SHARED_TOKEN=dev-secret \
PRISM_DATABASE_URL=sqlite+aiosqlite:///./prism.sqlite3 \
.venv/bin/uvicorn prism_challenge.app:app --host 0.0.0.0 --port 8000

Source: docs/operators.md:110-116.

Health checks

curl http://localhost:8000/health
curl http://localhost:8000/version

Internal weights require the shared token:

curl -H "Authorization: Bearer dev-secret" \
  -H "X-Platform-Challenge-Slug: prism" \
  http://localhost:8000/internal/v1/get_weights

Source: docs/operators.md:124-137.

Deployment

In a deployment, PRISM registers as a challenge image reached by the master over the internal challenge network. Public miner traffic goes through the proxy, which verifies signatures and forwards to PRISM. Weights are exposed only via get_weights and are always dry-run. Source: docs/operators.md:118-122.

Troubleshooting

Symptom	Likely cause
`invalid internal token`	Shared token mismatch between the master and PRISM
submission rejected before container	Static sandbox, two-script contract, param cap, distributed contract, or LLM hard-gate reject
submission held	LLM review quarantine (transient error or ambiguous verdict)
evaluation failed	Broker, image, GPU, timeout, missing locked data, or container error
empty weights	No completed, scored submissions yet
`missing_locked_data`	The read-only FineWeb-Edu train mount is absent or empty on the GPU node

Source: docs/operators.md:139-148.

Overview

Agent Challenge

PRISM

Additional challenges

Build a challenge

PRISM challenge operators

Installation

Local validation

Required runtime configuration

Docker broker configuration

Locked FineWeb-Edu data plane

Compute budget

LLM hard gate configuration

Multi-GPU static contract

Duplicate review

Running locally

Health checks

Deployment

Troubleshooting

​Installation

​Local validation

​Required runtime configuration

​Docker broker configuration

​Locked FineWeb-Edu data plane

​Compute budget

​LLM hard gate configuration

​Multi-GPU static contract

​Duplicate review

​Running locally

​Health checks

​Deployment

​Troubleshooting

Installation

Local validation

Required runtime configuration

Docker broker configuration

Locked FineWeb-Edu data plane

Compute budget

LLM hard gate configuration

Multi-GPU static contract

Duplicate review

Running locally

Health checks

Deployment

Troubleshooting