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.

Once the Pareto frontier has been computed, Bitrecs V2 applies winner-takes-all (WTA) scoring to determine which frontier miner receives onchain weight. The algorithm evaluates every possible combination of task environments (every non-empty subset), finds a winner for each subset, accumulates points by subset size, and converts the final point totals into normalized weights via softmax. Only one miner — the one with the highest resulting weight — has set_weights called in their favour.

​
Algorithm overview

1

Enumerate subsets

For every non-empty subset S of the active task environments — from singletons up to the full set — the engine identifies which frontier miner wins on S.
2

Score subsets by size

Each subset contributes points equal to its size (the default LINEAR scheme). A subset of 3 environments awards 3 points to its winner. Larger subsets therefore matter more.
# scoring/wta.py
match subset_weight_scheme:
    case SubsetWeightScheme.LINEAR:
        subset_weight = float(subset_size)
    case SubsetWeightScheme.EXPONENTIAL:
        subset_weight = float(2 ** (subset_size - 1))
    case SubsetWeightScheme.EQUAL:
        subset_weight = 1.0
3

Find each subset winner

For each subset, eligible miners are ranked by their total score across environments in the subset. The top-ranked miner wins if they clearly beat the runner-up’s threshold on a majority of environments. Otherwise the tie is broken by first-commit block (earlier = wins).
4

Accumulate and convert to weights

Each miner’s subset points are summed. The resulting scores are passed through scores_to_weights, which applies softmax normalization to produce weights that sum to 1.

​
Threshold computation

A “clear win” over the runner-up is determined by compute_miner_thresholds, which computes per-miner, per-environment thresholds based on a Z-score confidence band: 
# scoring/threshold.py
def calculate_threshold(
    prior_score: float,
    sample_count: int,
    z_score: float = DEFAULT_Z_SCORE,   # 1.5
    min_gap: float = MIN_THRESHOLD_GAP, # 0.02
    max_gap: float = MAX_THRESHOLD_GAP, # 0.08
) -> float:
    p = max(0.01, min(0.99, prior_score))
    se = math.sqrt(p * (1.0 - p) / sample_count)
    gap = z_score * se
    gap = max(gap, min_gap)
    gap = min(gap, max_gap)
    return min(prior_score + gap, 1.0)
ConstantValueMeaning
DEFAULT_Z_SCORE1.5Confidence multiplier applied to the standard error
MIN_THRESHOLD_GAP0.02A later miner must beat the earlier by at least 2 pp
MAX_THRESHOLD_GAP0.08Gap is capped at 8 pp regardless of sample size
More evaluation samples reduce the standard error, which shrinks the threshold gap (down to the 2 pp floor). A well-sampled subnet makes it easier for a genuinely superior miner to claim a win.

​
Subset winner selection

find_subset_winner_score_first implements the per-subset decision: 
# scoring/wta.py
def find_subset_winner_score_first(
    miner_scores: MinerScores,
    miner_thresholds: MinerThresholds,
    miner_first_blocks: MinerFirstBlocks,
    subset: tuple[EnvironmentId, ...],
) -> MinerUID | None:
    eligible = [
        u for u in miner_scores
        if any(miner_scores[u].get(e, 0.0) > 0.0 for e in subset)
    ]
    if not eligible:
        return None
    if len(eligible) == 1:
        return eligible[0]

    def subset_score(uid):
        return sum(miner_scores[uid].get(e, 0.0) for e in subset)

    # Sort by score descending first, then block ascending as secondary
    ranked = sorted(
        eligible,
        key=lambda u: (-subset_score(u), miner_first_blocks.get(u, float("inf")))
    )

    leader    = ranked[0]
    runner_up = ranked[1]

    n_subset = len(subset)
    majority = (n_subset + 1) // 2  # ceil(n/2)

    wins_over_runner_up = sum(
        1 for env in subset
        if miner_scores[leader].get(env, 0.0)
        > miner_thresholds[runner_up].get(env, miner_scores[runner_up].get(env, 0.0) + MIN_THRESHOLD_GAP)
    )
    leader_is_clear = wins_over_runner_up >= majority

    if leader_is_clear:
        return leader

    # Statistical tie: first committed block decides
    return min(
        [leader, runner_up],
        key=lambda u: (miner_first_blocks.get(u, float("inf")), u)
    )
First-commit advantage only applies when the score difference between leader and runner-up is within the statistical threshold band. A miner with a clear score advantage always wins regardless of block order.

​
Converting scores to weights

Once all subset points are accumulated, scores_to_weights applies temperature-scaled softmax: 
# scoring/wta.py
def scores_to_weights(
    scores: dict[MinerUID, float],
    temperature: float = 1.0,
    min_weight: float = 0.0,
) -> dict[MinerUID, float]:
    uids = list(scores.keys())
    values = np.array([scores[uid] for uid in uids])

    if np.all(values == 0):
        n = len(uids)
        return {uid: 1.0 / n for uid in uids}

    # Softmax with temperature (shifted for numerical stability)
    values_shifted = values - np.max(values)
    exp_values = np.exp(values_shifted / temperature)
    weights = exp_values / exp_values.sum()

    return {uid: float(w) for uid, w in zip(uids, weights, strict=True)}
The miner with the highest weight is the WTA winner: 
# scoring/engine.py
weight_receiving_uid = max(weights, key=weights.get)

​
Inspecting WTA results

The /scoring/wta endpoint exposes the current subset scores, weights, and per-subset winners for the active evaluation set: 
GET /scoring/wta
The response includes weights, subset_scores, miner_thresholds, miner_blocks, and a sample of subset_winners keyed by subset tuple string.