Michael Darius Eastwood · Independent AI alignment researcher
Published
Graduated Autonomy is the Eden Protocol principle that AI systems should receive increasing autonomy as they demonstrate trustworthiness, analogous to how a child is given increasing independence over time. It is contrasted with binary approaches (fully constrained or fully autonomous). The design intent is that trust is earned through visible track record, and revocable if the system fails a downstream check.
The framework treats autonomy as a granted resource rather than a default state. Each increment of decision authority requires an antecedent demonstration of alignment behaviour that survives adversarial pressure, most importantly the Monitoring Removal Test, where behavioural drift between observed and unobserved conditions must remain close to zero. Autonomy withdrawal is symmetric: any observed drift, coupling degradation, or failure of an ethical loop reduces the permitted decision surface until the system re-establishes evidence of stable alignment. The principle deliberately imports vocabulary from developmental parenting and from safety-critical certification, where the presumption is that novel actors begin with narrow, high-supervision authority and expand only under measured competence.
Graduated Autonomy is specified in the Eden Engineering paper as part of the governance architecture that complements the Chokepoint Strategy and the HARI Treaty. It is the operational corollary of the load-bearing safety commitment: only systems whose safety is structurally verified receive increased autonomy.
The engineering context matters. The Eden Engineering paper argues that any alignment strategy with an alignment-scaling exponent below the capability-scaling exponent is guaranteed to fail at sufficient recursive depth. Graduated Autonomy is the governance layer that mirrors that mathematics: capability tier is bounded by verified alignment tier, and increased autonomy is contingent on evidence that both scale together. In the paper's four-tier deployment ladder the top tier is reserved for systems that pass the Monitoring Removal Test at depth and preserve coupling under adversarial suppression prompts.
The general direction (staged autonomy contingent on demonstrated safety) is consistent with the broader AI-safety literature on levels of autonomy and with regulatory approaches in adjacent industries (aviation, nuclear). No independent programme has proposed a specific Graduated Autonomy schedule matched to the Eden Protocol's Monitoring Removal Test or its coupling parameter measurement.
Governance direction, not deployed policy. The Eden Engineering paper specifies the principle but does not define quantitative thresholds. The programme's empirical work concentrates on the substrate mechanisms (loops, entangled loss, load-bearing safety); Graduated Autonomy is the framework within which those mechanisms would be operationalised at deployment. The Graduated Autonomy schedule inherits every open question from the underlying measurements: alignment-scaling exponents have only been measured across six frontier models in a single blind evaluation, the Monitoring Removal Test protocol is defined but not yet run at depth, and the coupling parameter link is a hypothesis about how ethics can be tied to the beta scaling term rather than a validated engineering procedure. Deployment thresholds therefore remain deliberately unquantified until the substrate research produces the numbers that a real autonomy schedule would depend on.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.