Runtime governance for adaptive AI
A control layer that sits between an AI system's decision and its action. It grants each action only the authority that action has earned at that moment, cannot be bypassed before the actuator, and writes a tamper-evident record of what it allowed.
It is distinct from model safety, which acts at training time, and from binary guardrails, which only allow or block. Runtime governance modulates authority continuously and leaves per-decision evidence a regulator can read.
Assured Autonomy Governor (AAG)
Wayfinder's runtime governor for cyber-physical actuation. It reads the divergence between what a system expected and what it observed, maps that to a bounded trust score, and from there to an authority ceiling that can only fall, enforced before the actuator.
Every governed command leaves a hash-chained record. It sits above the control API and below the autonomy.
Assured Predictive Governor (APG)
The forward-looking layer of the governor. A Koopman-operator kernel projects where the system's trajectory is heading, so authority can be pulled back before drift reaches the actuator.
This is why a more confident decision can be granted less authority: the forward look caught a divergence the current step had not yet surfaced. Prediction here serves the governance guarantee; it is not a standalone forecast product.
Assured Learning Governor (ALG)
Wayfinder's runtime governor for learning updates. Instead of triggering a retrain when a model drifts, it gates or rolls back the learning event itself, through multi-axis divergence, phase-based authority modulation, and a two-layer rollback, with a cryptographic audit chain.
It governs the update, not just the output.
NeuroPRIN
The neural trust-inference kernel inside the governors. A multi-head model infers a bounded trust score from expected-versus-observed divergence, and that score sets how much authority the next action is granted.
NeuroPRIN runs bit-identically across the Python, C++, and embedded builds, so the trust it computes is the same on a laptop and on the target device.
Tamper-evident AI audit record
The per-decision output of the governor. Each governed action produces one record of the action, the authority it was granted, and the resulting safety state, cryptographically chained to the one before it so any change is detectable.
A model, an auditor, and a regulator read it from the same line. It exports to formats regulators already use, including the NAIC Annual Statement, the Solvency II SFCR, and AM Best BCAR.
