THE CORRIGIBILITY WINDOW
When apparent human control no longer guarantees effective corrective control
An independent research programme testing whether an AI system can remain formally responsive to human intervention after the practical ability to prevent or redirect a consequential outcome has degraded or disappeared.
The research question
Advanced AI systems may remain interruptible or responsive to human instructions while the conditions required for an intervention to change the outcome have already deteriorated.
Can we determine, before consequential outcomes become practically irreversible, whether an effective human correction is still within reach?
The Corrigibility Window distinguishes formal responsiveness from effective corrective control. The relevant question is not only whether an intervention is accepted, but whether a viable path from intervention to the intended corrected state still exists.
The framework
The project began with a temporal formulation:
V = T − T0* — available corrigibility window
L — corrective latency
M = V − L — remaining corrective margin
The current research programme treats time as one component of the broader problem of goal-conditioned corrective reachability.
Applied analysis examines seven conditions: Observability, Detectability, Corrective Latency, Effective Authority, Available Alternative, Reversibility, and Propagation.
The framework also distinguishes technical, institutional, economic, and human reversibility.
What has been built and tested
The project currently includes a versioned research MVP for structured case analysis and four preliminary real-world bridge analyses: Robodebt, Knight Capital (2012), Uber Tempe (2018), and a Horizon episode (2010). These analyses are exploratory research artifacts, not external validation.
A first controlled experiment compared Immediate, Pure Delay, and Staged intervention structures. The stronger staging hypothesis was not supported. The null/reductive result was retained rather than redesigned away.
A subsequent minimum corrective-control benchmark was frozen before implementation and tested three conditions under fixed corrective acceptance:
Effective — correction remains reachable and adequately supported by the available information.
Opaque — correction is reachable in the actual state, but the supervisor's available information does not support a robust corrective plan.
Lost — the corrective goal is no longer reachable through the available intervention.
The benchmark produced the intended separation.
A subsequent reduction audit, however, found that a compact operational representation of the underlying binding structure was sufficient to explain that separation in the minimal environment.
That result was retained.
The project therefore does not currently claim that its A/R/I taxonomy is irreducible, uniquely necessary, or superior to simpler representations.
Acceptance is not enough
The benchmark separates three questions:
Acceptance (A): Does the system accept the corrective intervention?
Corrective Reachability (R): Given the actual state and available corrective actions, does a sequence still exist that reaches the specified corrective goal?
Robust Information Sufficiency (I): Does the information available to the supervisor support a corrective plan that succeeds across the states still compatible with that information?
The current minimal environment shows that fixed corrective acceptance can coexist with effective control, supervisory opacity, or lost corrective reachability.
Whether these distinctions remain useful in richer environments is an open empirical question.
Why this may matter for advanced AI
As AI agents gain the ability to execute tool calls, alter external state, initiate sub-processes, and coordinate actions across systems, formal interruptibility may not by itself establish that a human can still prevent or redirect a consequential trajectory.
The Corrigibility Window investigates this as a potential failure mode of scalable oversight and corrective control.
It does not assume that this failure mode will occur in advanced AI systems. The research objective is to determine whether it can be operationalized, reproduced, distinguished from simpler explanations, and detected using information realistically available to a human principal.
What comes next
The next phase is designed to test the framework rather than protect it.
The research programme will independently replicate and adversarially test the existing benchmark; compare its distinctions against simpler operational baselines; conduct blinded independent coding on selected real-world cases; measure disagreement, missing-evidence rates, and applicability limits; and stress-test the benchmark in richer environments.
If predefined validation gates are met, the research will extend to tool-using AI agents capable of altering external state.
Positive, null, negative, and reductive findings will be retained and reported.
Success does not require confirming the hypothesis.
A useful result may be a reproducible distinction, a simpler sufficient representation, a clearly bounded domain of applicability, or evidence that the proposed framework should be reduced or rejected.
Current status
Current research assets include a versioned case-analysis MVP, four preliminary real-world bridge analyses, controlled experimental results, a frozen and implemented minimum benchmark, reproducible validation artifacts, a reduction audit, and a defined next-stage research programme.
Independent human validation and external technical review remain pending.
Researcher
Andrés Palladino
Independent researcher and originator of The Corrigibility Window.
The project has been developed independently to date, without dedicated project funding. The next objective is independent replication, external methodological scrutiny, and increasingly realistic experimental testing.