aether_quant.moe¶
Mixture-of-experts gating.
Mixture-of-experts gating.
ExpertGateWeight
dataclass
¶
One expert's weight in the gated blend, with its raw score.
GatingDecision
dataclass
¶
The gated probability of an up move, the baseline and expert inputs, and each expert's weight.
build_gating_decision ¶
build_gating_decision(regime: dict[str, Any], expert_training_metrics: dict[str, Any], expert_probabilities: dict[str, float | None], baseline_probability_up: float | None = None, baseline_weight: float = 0.25, gating_model: dict[str, Any] | None = None, gating_feature_schema: dict[str, Any] | None = None, expert_magnitudes: dict[str, float | None] | None = None, expert_volatilities: dict[str, float | None] | None = None, baseline_magnitude: float | None = None, baseline_volatility: float | None = None, sequence_prediction: dict[str, Any] | None = None, sequence_weight: float = 0.0, excluded_experts: tuple[str, ...] | frozenset[str] = ()) -> GatingDecision
excluded_experts (V6.4.0, phase_v2.gating_network.disabled_experts, default none): experts treated as having no
probability, so the others renormalise - the expert-ablation switch (lead L8).
build_gating_model_features ¶
build_gating_model_features(regime: dict[str, Any], baseline_probability_up: float | None, weights: list[ExpertGateWeight]) -> list[float]
Pure feature-vector builder for the optional learned gating model -
same fixed order as GATING_MODEL_FEATURE_KEYS. weights must contain
exactly one ExpertGateWeight per EXPERT_NAMES entry, which
build_gating_decision() always guarantees.