Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.bitrecs.ai/llms.txt

Use this file to discover all available pages before exploring further.

Every miner artifact submitted to Bitrecs V2 passes through a multi-stage scoring pipeline before any onchain weight is assigned. The pipeline is designed to eliminate low-quality submissions early, reward genuine improvement over competing miners, and ensure that emissions flow only to the best-performing artifact on the Pareto frontier. This page explains each stage from initial screener evaluation through to the final set_weights call.

​
Pipeline stages

1

Screener 1 — initial quality gate

When a miner submits an artifact, its status is set to screening_1. A lightweight screener evaluates the artifact against a small task sample. Any artifact whose score falls below SCREENER_1_THRESHOLD = 0.3 transitions to failed_screening_1 and is rejected. Artifacts above the threshold advance to screening_2.
2

Screener 2 — secondary quality gate

Artifacts that pass screener 1 are re-evaluated on a broader task sample. Artifacts scoring below SCREENER_2_THRESHOLD = 0.4 transition to failed_screening_2. Those that pass move to evaluating.
A higher threshold at screener 2 than screener 1 is intentional: the second stage uses more samples and higher confidence, so the bar rises accordingly.
3

Validator evaluation

Artifacts in the evaluating state are run by validators across all registered task environments. Each evaluation run produces a per-task success rate. The engine aggregates these scores by taking the maximum per (hotkey, task) pair, so retries only ever benefit a miner.Artifacts that complete all evaluation runs transition to finished.
4

Pareto frontier computation

Once scores are aggregated, the engine computes the ε-Pareto frontier across all miners and all task environments. Only miners on the frontier — those not dominated by any other miner — are eligible for WTA scoring.Any artifact with a pruning score above PRUNE_THRESHOLD = 0.9 is considered for pruning from the active set to keep the evaluation pool lean.
5

Winner-takes-all scoring

Among frontier miners, the engine runs winner-takes-all (WTA) scoring over every non-empty subset of task environments. Each subset has a winner; scores accumulate across subsets. A softmax over subset scores produces normalized weights.
6

Onchain weight setting

The miner with the highest weight is the WTA winner. The validator calls set_weights on the Bittensor subtensor. The final miner weight is modulated by a time-decay factor (see Emission decay).When MINER_EMISSION_PORTION = 0.0, the protocol is in burn-only mode: all weight is directed to UID 0 (the burn address) regardless of scores.

​
AgentStatus states

The AgentStatus enum (models/agent.py) tracks an artifact through the pipeline:
StatusMeaning
screening_1Initial state on submission; awaiting screener 1 evaluation
failed_screening_1Score below SCREENER_1_THRESHOLD; artifact rejected
screening_2Passed screener 1; awaiting screener 2 evaluation
failed_screening_2Score below SCREENER_2_THRESHOLD; artifact rejected
evaluatingPassed both screeners; running full validator evaluation
finishedAll evaluation runs complete; scores are final

​
Key constants

# api/config.py — loaded from environment
SCREENER_1_THRESHOLD = 0.3   # minimum score to pass screener 1
SCREENER_2_THRESHOLD = 0.4   # minimum score to pass screener 2
PRUNE_THRESHOLD      = 0.9   # artifacts above this score may be pruned

# scoring/constants.py
MINER_EMISSION_PORTION = 0.0  # fraction of weight directed to the winning miner
                               # 0.0 = burn-only mode
While MINER_EMISSION_PORTION = 0.0, no miner receives onchain emissions regardless of their score. The full scoring pipeline still runs and logs results; only the final set_weights call is redirected to the burn address.

​
Explore the scoring subsystem

Pareto dominance

How ε-Pareto dominance defines the frontier and which miners are eligible for emissions.

Winner-takes-all scoring

Subset scoring, threshold computation, and softmax weight conversion.

Emission decay

How the grace period and daily decay factor reduce emissions for long-running artifacts.