Skip to main content

Module probe_runner

Module probe_runner 

Source
Expand description

ProbeRunner — the closed loop between the steering primitive (crate::inference::steering) and the anytime-valid structure-evidence ledger (gam_terms::inference::structure_evidence).

Both halves are implemented and tested in isolation; nothing wired them into a runnable experiment loop. This is that bridge. The evidence module decides WHICH claim to interrogate and HOW MUCH a probe should move the e-process (plan_probe_for_contested_claim); the steering module turns a chosen latent intervention into the on-manifold activation delta with its dosimetry and validity radius (steer_delta). The runner picks the contested claim, asks the planner for the experiment, realizes it through the steering primitive, and feeds the realized dose back into the ledger as anytime-valid evidence.

§The discrimination coordinate

For a contested claim about atom k, the two hypotheses are “this atom carries the steering move along its learned surface” (the alternative) versus “it does not” (the null). The steering primitive measures, in nats of output-Fisher KL, exactly how much behavioral effect a latent move along atom k actually delivers — its predicted_nats dose. That dose IS the expected per-observation log-growth of the deciding e-process under the alternative (the module docs’ “the SAME quadratic form the steering dosimetry already computes, repurposed”). So each candidate latent move becomes a one-dimensional CandidateProbe whose hypothesis disagreement, read through the identity Fisher, reproduces the realized dose: ½ (μ₁ − μ₀)ᵀ F (μ₁ − μ₀) = predicted_nats with μ₀ = 0, μ₁ = √(2·predicted_nats), F = [[1]]. No fabricated metric — the real steering dose flows through the real planner.

Structs§

ProbeRunner
The closed-loop probe runner over one fitted SAE-manifold term and its per-row output-Fisher metric.
RealizedProbe
A planned probe carried alongside its realized steering intervention: the experiment-design output (plan), the on-manifold activation delta and dosimetry the steering primitive produced for it (steer), and the realized behavioral dose in nats once it has been measured (realized_nats, None until ProbeRunner::absorb banks it).