Subsystem: autotune — offline binding-spec inference

Infers a binding spec from raw time series so a domain can be onboarded from data rather than hand-authored. 14 files, ~5.9k LOC. Offline; not in the step loop.

Inputs

Multichannel signals (n_channels, n_samples) + sampling rate; optionally a graph or an event log.

Outputs

AutoTuneResult(omegas, knm, alpha, n_layers, dominant_freqs, K_c_estimate); discovered symbolic equations (SINDy); BindingSpec proposals ready for the CLI or server.

Processing model

identify_binding_spec chains: phase extraction → frequency identification by exact Dynamic Mode Decomposition → least-squares coupling estimation → critical- coupling prior. PhaseSINDy performs sparse symbolic regression over a trigonometric library. RL-style policy search, learner generators (PPO/SAC/hybrid), knob attribution, and a candidate-safety certificate round out the lane.

Backends

SINDy has a Rust path; coupling estimation is Python-only (least-squares).

Wiring

propose_binding_from_* produces a BindingSpec consumed by load_binding_spec and then the normal pipeline. SINDy output is for inspection.

Scope boundaries

The candidate-safety certificate generates a report but does not block a proposal (no enforcement gate). It still validates replay evidence strictly: empty observation windows, malformed replay states, and non-finite barrier margins fail before a certificate is emitted. Policy-search and learner outputs are proposals, not live actuation.