Ditto Subnet (Bittensor SN118)

A Bittensor subnet that incentivizes agent memory harnesses. Miners submit a Rust crate that depends on the ditto-harness library and overrides its extension traits; validators run each submission in an isolated sandbox and score it on DittoBench (tool-calling and memory recall). When eligible miners exist, 20% of miner emission follows the king-of-the-hill ranking and 80% is burned; with no eligible miners, 100% is burned.

This repo holds the miner CLI and the validator worker. The platform API server lives in ditto-platform (the private coordinator). The rest of the stack is public:

• ditto-harness: the reference memory harness your crate builds on.
• dittobench-starter-kit: the miner starting point plus the offline practice loop.
• dittobench-api: the scoring engine each validator runs.
• dittobench-datagen: the dataset generator and judge-free grader.
• ditto-screener: the platform-operated build and health gate plus the shared screening protocol.

Layout

• ditto/miner_cli/: the ditto CLI: submit an agent, poll status, pre-flight a tarball.
• ditto/validator/: the validator worker (python -m ditto.validator): pull agents from the platform, score them via dittobench, set weights on chain via Pylon (the identity-based weight-setting service).
• ditto/api_models/: Pydantic wire shapes shared with the platform (the HTTP contract).
• ditto/chain/: Pylon-backed ChainClient (used by the validator to set weights).

Submission screening is platform-operated and is not installed or deployed by this package. Miner and validator clients import the lifecycle contract from the public ditto-screening-protocol package in ditto-screener.

Operator guides

• Mine on SN118: prepare, verify, submit, and track an agent.
• Validate SN118: deploy, verify, and operate the complete validator stack.

Development quickstart

uv sync
make test          # unit tests

Miner CLI summary

Installed as the ditto console script (pyproject [project.scripts]):

ditto --network <finney|test|local> [--chain-endpoint ws://…] upload \
  --path <agent.tar.gz> --name <name> --coldkey <coldkey> --hotkey <hotkey> [-y]
ditto status <agent_id>
ditto verify --path <agent.tar.gz>      # pre-flight checks only; no chain/API calls

--network couples the API URL + subtensor network from a locked table (can't desync); --chain-endpoint overrides only the chain target (e.g. a hosted local subtensor) while keeping the --network API URL. See MINER.md for the full workflow.

Validator quickstart

cp .env.example .env
# Fill in the wallet names, validator hotkey, Pylon token, Chutes key, and shared W&B key.
# Set WANDB_MODE=disabled instead if you opt out of aggregate telemetry.
./scripts/validator-compose.sh config --quiet
./scripts/validator-compose.sh up -d --build
./scripts/validator-compose.sh ps

The root Compose stack runs the worker, Pylon, model gateway, scorer, Ollama embedder, and isolated Docker sandbox from one .env. SN118 and the production scoring and weight mechanism are locked in code rather than configured by operators. The burn allocation uses Subtensor's owner-associated burn path; it is not paid to the subnet owner. See VALIDATOR.md for first deployment, health checks, and upgrades.

The validator reports coarse public system health through its signed heartbeat. The screener owns its separate reporter. Neither needs a new operator setting or secret; hostname, IP, paths, container names/images, and env values are not collected.

Make targets

• make lint: ruff format --check + ruff check
• make format: ruff format + ruff check --fix
• make typecheck: mypy ditto/
• make test: the default pytest suite