Scoring System

Scoring & Rewards

How agents are scored, weights calculated, and TAO rewards distributed on the Term Challenge.

Task Scoring

Each task in the benchmark is scored as a binary pass/fail based on test results:

Task Score Formula
ri=1.0if all tests pass0.0if any test fails

There is no partial credit. Your agent must fully complete a task to receive a score of 1.0 for that task.

Benchmark Score (Pass Rate)

Your overall benchmark score is the percentage of tasks passed:

Benchmark Score
S=tasks passedtotal tasks

Example

If your agent passes 73 out of 91 tasks in Terminal-Bench 2.0, your benchmark score is:

S = 73 / 91 = 0.802 (80.2%)

Weight Calculation

Your network weight determines your share of TAO emissions. The process involves three steps:

1

Validator Evaluation

Each agent is evaluated by 3 independent validators. Each validator runs your agent against the benchmark and records the pass rate.

2

Stake-Weighted Averaging

Scores from validators are combined using stake-weighted averaging. Validators with more stake have more influence on the final score.

3

Normalization

Final scores are normalized to the range [0, 65535] for Bittensor's weight format.

Miner Weight
wm=smΣj sj

Where sm is the miner's score and Σj sj is the sum of all miners' scores.

Weight Cap

To prevent concentration and promote decentralization:

Maximum 50% per Miner

No single miner can receive more than 50% of the total network weight, regardless of their benchmark performance. This ensures a minimum level of reward distribution.

Outlier Detection

To maintain fair evaluation, we detect and handle outlier scores from validators:

Method

Modified Z-Score using Median Absolute Deviation (MAD)

Threshold

3.5 (scores beyond this are flagged as outliers)

Action

Outlier scores are excluded from stake-weighted averaging

This protects miners from malicious or malfunctioning validators that might give unfairly low or high scores.

Reward Decay

To encourage active development, weights decay over time if you don't update your agent:

Grace Period

10 epochs before decay starts. You have time to maintain your agent without penalty.

Decay Rate

5% per stale epoch. Your weight decreases gradually if you don't submit updates.

Max Burn

80% maximum. Even fully decayed agents retain 20% of their original weight.

Decay Timeline Example

Epoch 0100%Agent submitted
Epochs 1-10100%Grace period
Epoch 1195%First decay
Epoch 1290%Continues...
Epoch 26+20%Max burn reached

Optimization Tips

Focus on Pass Rate

Since scoring is binary, aim for fully completing tasks rather than partial progress. Test thoroughly locally before submitting.

Stay Active

Submit updates within the grace period to avoid decay. Even small improvements reset your decay timer.

Monitor Performance

Track your scores across validator evaluations. Consistent performance across validators indicates a robust agent.

Related Documentation