pub struct Test {}Expand description
The randomised question-count test of the learned policy (select/policy.rs choose; docs/delivery.md
“The question-count test”), all-time as the summary counts it (select/summary.rs Experiment), never
recounted from the log. On a live check, of the k lessons below grade 3 the fixed rule delivers and
that may go free, s drawn uniformly in 0..=k are asked and the rest handed over: nothing is withheld.
Fields§
§risk: f64The share of the fixed rule’s value the gate may risk on it: the collective’s risk in
collectives.json (select.rs risk_of), else select/policy.rs ALPHA. 0: the test is off.
set: boolWhether the collective sets risk (false: the default ALPHA).
ran: u64Checks the test drew a count for (played or refused by the gate), and of them the control arm’s
(experiment_control: s is the fixed rule’s own count, so it routed as the fixed rule).
control: u64§refused: u64Of them, checks the gate refused it: the draw’s risk did not fit what the gate may still risk, so
the fixed rule decided (source gate).
sessions: u64The sessions with a check the test drew a count for.
kept: u64Of ran, the drawn checks the summary still holds, which the readout’s estimates read
(select/replay.rs TestReadout::kept): compaction drops the oldest checks, never the counts.
by_k: Vec<Draws>Per k, how often each s was drawn (draws[s], s from 0 to k) and the checks refused at that k.
due: boolThe pre-registered readout (select/replay.rs test_readout): due after TEST_CHECKS checks drawn
or TEST_SESSIONS sessions with one, and only then estimated, each with its session-bootstrap
95% interval: per number m ≥ 2 of the test’s questions asked, the test’s top lesson’s change rate
against asked alone (dilution, unit points changed), and asking a grade 2 against handing it
over (ask_free, unit points changed). Empty and None until due, or with no drawn check.
due_checks: u64§due_sessions: u64§dilution: Vec<Estimate>§ask_free: Option<Estimate>