Struct ProbeRunner 

pub struct ProbeRunner<'a> {
    pub term: &'a SaeManifoldTerm,
    pub metric: &'a RowMetric,
}

The closed-loop probe runner over one fitted SAE-manifold term and its per-row output-Fisher metric.

Fields

term: &'a SaeManifoldTerm

The fitted term whose atoms the probes steer (read only).

metric: &'a RowMetric

The per-row output-Fisher inner product the dose is measured through.

Implementations

impl<'a> ProbeRunner<'a>

pub fn design_next( &self, ledger: &StructureLedger, ) -> Result<RealizedProbe, String>

Design the next probe for the most contested claim in ledger.

Picks the contested claim with the LEAST evidence (the one furthest from the 1/α Ville threshold — the most in need of interrogation), reads the atom it concerns, and builds candidate steering moves along that atom’s latent axes from its fitted representative coordinate. Each candidate is realized through steer_delta so its actual output-Fisher dose is known; the candidates are handed to plan_probe_for_contested_claim, which selects the most discriminating one and converts the claim’s current evidence into a remaining budget. The selected candidate’s already-computed SteerPlan rides back in the result.

pub fn absorb( &self, ledger: &mut StructureLedger, probe: &RealizedProbe, realized_nats: f64, )

Absorb a realized probe outcome, updating the ledger’s evidence for the probe’s claim.

realized_nats is the dose the probe actually delivered when run (the observed output-Fisher KL of the steered response). Under the local Gaussian output model the alternative-vs-null log-likelihood ratio of one such observation is exactly that dose, so it routes straight into the claim’s e-process through StructureLedger::absorb_probe_outcome as log(alt) − log(null) = realized_nats − 0. The contract its docstring requires — both hypotheses’ densities frozen before the outcome — holds here: the steering plan (and thus both predictions) was fixed at design time, before any outcome existed.