Member Junction
    Preparing search index...

    The gate: the transcribed decision table (BeatsByMargin, GateInputsFromIntervals, DecideGate), the stored decision with its guards (DecideGuardedGate), member names, stored intervals, the reason in words, and the shared interval text format.

    Index

    Constructors

    Properties

    HeuristicPrefix: "heuristic:" = 'heuristic:'

    How a member name marks the unit's existing heuristic rule (heuristic:<name>).

    Methods

    • The two guards outside the transcribed table: the rarest holdout class must have at least minMinorityClassRows rows (any outcome that ships), and a binary model whose holdout AUC interval contains 0.5 (or is missing) never ships. Guards only ever downgrade to no-signal. Publish re-verification calls DecideGuardedGate with the stored guard inputs, not DecideGate.

      The AUC guard reads the interval at the level the trainer reported it (0.95); it does not consult the protocol's promotion.lowerBoundLevel. The rule is "contains 0.5": an interval wholly below 0.5 does not trigger it (such a model would also have had to beat the dummy's loss by the margin to reach this point).

      Parameters

      Returns GuardedGateDecision

    • The best simple member: the lowest holdout score, first in name order on a tie (the sidecar picks it the same way). modelVsBestSimple is the model's comparison with this member.

      Parameters

      • scores: Readonly<Record<string, number>>

      Returns string | null

    • The decision table, a line-for-line transcription of the reference pipeline's gate_decision and of the sidecar's app/gate.py: (a) the model ships when it beats the dummy and the best simple member (and is not itself a simple member), or when it beats the dummy and no simple member does; (b) otherwise the eligible simple member with the lowest score ships; (c) otherwise nothing ships. Do not edit this or either copy without running the gate parity tests.

      Parameters

      Returns GateDecision

    • The intervals the run stores, read from the trainer's paired comparisons (modelVsDummy, <member>VsDummy, modelVs:<member>, as the sidecar keys them). Returns the first missing key instead of guessing.

      The delivered member is the finalist's name when that name is a simple member's key, else null. Candidate names can never be member names (the sidecar refuses reserved names), so a live run stores null — including where the sidecar passes glm as the delivered member for an untuned GLM finalist. The decision is the same either way: that finalist's predictions are the GLM's, so modelVs:glm is 0 [0, 0] and the model cannot beat the best simple member.

      Parameters

      Returns IntervalsFromResult