TurboVision

TurboVision is Score's decentralized intelligence layer for live video and imagery. The network pairs expert models with a global community of validators and miners so raw footage becomes structured, decision-ready data in real time, with early deployments focused on professional sports.

Participate

• Validators keep the network honest by scoring submitted models on live data. Read VALIDATOR.md.
• Miners contribute models that solve Elements in the Manifest. Read MINER.md.

Common Setup

Bittensor Wallet

Install the CLI, create a coldkey and hotkey, then copy the hotkey folder and public coldkey (coldkeypub.txt) onto every host that will run TurboVision.

pip install bittensor-cli
btcli wallet new_coldkey --n_words 24 --wallet.name my-wallet
btcli wallet new_hotkey --wallet.name my-wallet --n_words 24 --wallet.hotkey my-hotkey
cp env.example .env

Set these in .env:

• BITTENSOR_WALLET_COLD
• BITTENSOR_WALLET_HOT
• CHUTES_API_KEY
• HUGGINGFACE_USERNAME
• HUGGINGFACE_API_KEY

Chutes Access

Upgrade to a developer-enabled account on chutes.ai, install the CLI, register, and mint an API key. Store it as CHUTES_API_KEY.

pip install -U chutes
chutes register

Hugging Face Credentials

Create (or reuse) a Hugging Face account, generate a token with write access, and set HUGGINGFACE_USERNAME and HUGGINGFACE_API_KEY.

ScoreVision CLI

Install the CLI with uv, then sync dependencies and verify the binary:

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv && source .venv/bin/activate
uv sync
sv

Runner

The runner executes scoring jobs per Element on a fixed block cadence. It fetches a challenge, builds ground truth (real or pseudo), scores miners, and emits results to R2.

sv -vv runner

Flow (per Element run):

1. Load the active Manifest.
2. Pull a challenge for the current element_id.
3. Determine the window ID and start block for timing metadata.
4. Build ground truth:
   • If elements[].ground_truth=true, fetch real GT from the API.
   • Otherwise generate pseudo-GT locally (SAM3).
5. Call eligible miners (from on-chain registry) and score outputs.
6. Emit a shard to R2 with evaluation payload + metadata.

Scheduling:

• Runner keeps per-element timers based on elements[].window_block (or tempo).
• If an Element does not define a cadence, it uses SV_DEFAULT_ELEMENT_TEMPO_BLOCKS (default 300).
• When the manifest changes, runner rebuilds its per-element schedule.

Quality gates:

• Pseudo-GT is retried until enough frames meet bbox thresholds.
• Tune retries with SV_PGT_MAX_BBOX_RETRIES and SV_PGT_MAX_QUALITY_RETRIES.

Validator

The validator aggregates recent scores per Element, chooses winners, and submits weights on-chain via the signer service.

sv -vv validate --manifest-path path/to/manifest.yml

Flow (per window):

1. Load the active Manifest for the current block.
2. For each Element:
   • Determine how far back to look using elements[].eval_window (days).
   • Pull recent scores for that Element and window.
   • Select the winner and build a per-element weight share.
3. Normalize total weights to 1.0 (add fallback weight if needed).
4. Submit uids + weights to the signer.
5. Optionally snapshot winners to R2.

Behavior knobs:

• elements[].eval_window controls lookback in days (converted to blocks).
• elements[].weight controls the Element share of final weights.
• Hotkey blacklist is loaded from manifest_update/blacklist (one hotkey per line, # for comments).
• Startup commit: SCOREVISION_COMMIT_VALIDATOR_ON_START=1 publishes the validator index.
• Winners snapshots are emitted on every successful set_weights (central validator only).

Community & Support

• Score Website
• File issues or ideas in this repo.