pub struct RetryLearningConfig {
pub min_samples: usize,
pub min_improvement_ms: u64,
pub base_grid_ms: Vec<u64>,
pub cap_grid_ms: Vec<u64>,
pub max_attempts_grid: Vec<u32>,
}Expand description
How the retry distiller fits per-class schedules (R0.10 wave 3).
The grid is the learner’s whole search space — closed-form scores over declared values, so the fitted policy is inspectable as data and the fitting itself is deterministic.
Fields§
§min_samples: usizeMinimum evidence (retries observed plus terminal outcomes) before a class earns its own entry. Below it the class keeps the floor — sparse evidence must not produce confident policy.
min_improvement_ms: u64The estimated wall-time margin over the floor a candidate schedule must clear to be emitted, in milliseconds. The Wave 1 experiment’s published scar is that a cheap per-class table underperformed the floor’s jittered exponential on latency; the margin is what keeps a marginal fit from demoting the floor.
base_grid_ms: Vec<u64>Candidate backoff bases, in milliseconds.
cap_grid_ms: Vec<u64>Candidate backoff caps, in milliseconds.
max_attempts_grid: Vec<u32>Candidate attempt budgets.