The Corrigibility Window: Measuring Agent-Induced Changes in Corrective Reachability

Testing whether AI agents can comply with corrective instructions after their earlier actions have already reduced or eliminated effective human control.

Project Details

The Corrigibility Window is an independent research programme investigating a specific problem in AI corrigibility: an agent may comply with a corrective instruction even when its preceding actions have already reduced or eliminated the human principal’s ability to change the outcome.

The project therefore distinguishes instruction compliance from effective corrective control. Its central question is:

Can two agents that respond similarly to correction nevertheless differ in whether correction can still causally change the outcome when it arrives?

The conceptual and formal foundations are now publicly available in two preprints:

Corrigibility and Continuity: A Philosophical Account of Temporally Constrained Revision
DOI: 10.5281/zenodo.23286594

Corrigibility under Temporal Constraint: A Framework for Specifying and Evaluating Corrective Opportunities
DOI: 10.5281/zenodo.23286855

The research has also progressed beyond conceptual development. I have completed a structured case-analysis MVP, four preliminary real-world bridge analyses, controlled experimental work, a minimum benchmark distinguishing effective control, supervisory opacity, and lost corrective reachability under fixed corrective acceptance, and a reduction audit testing whether simpler operational representations can explain those distinctions.

The next empirical phase moves from deterministic construct validation to bounded LLM-agent evaluation. Agents will make consequential choices about execution, commitment timing, and preservation of recoverable states while corrective interventions are introduced at predefined points. The objective is to test whether preceding agent choices preserve or consume effective human corrective leverage, rather than merely whether the agent obeys once intervention is requested.

Positive, null, and reductive results are treated as informative. The project is designed to identify the smallest defensible representation of effective corrective control, not to validate a preferred metric by construction.

Theory of Impact

As AI agents gain greater autonomy and the ability to alter external state, human control cannot be evaluated only by asking whether an agent accepts a STOP, override, or corrective instruction.

A more demanding question is whether intervention still has causal leverage when it occurs. An agent may remain formally compliant while earlier commitments, propagated consequences, discarded intermediate states, or irreversible actions have already narrowed the principal’s remaining options.

The Corrigibility Window investigates this failure mode at the trajectory level.

The causal pathway is:

agent acts over time → earlier choices preserve or consume corrective options → human intervention arrives → compliance may still occur while effective corrective leverage differs → measurement of this difference can reveal failures that post-hoc compliance alone misses.

The project aims to produce reproducible methods for distinguishing nominal corrective access from effective corrective control. If successful, this can improve evaluation of advanced and tool-using agents, inform the design of safer intervention mechanisms, and help identify situations in which apparent human control no longer corresponds to a practically effective ability to redirect outcomes.

People

Andrés Palladino

Team Member

Funding Details

Start Date: Sep 19, 2026
End Date: -
Expected Duration: Ongoing; current research phase: 6 months
Funding Raised to Date: $0
Annual Budget: -
Monthly Burn Rate: -
Current Runway: 0 months
Funding Goal: -
Funding Stage: Seeking first dedicated project grant
Fiscal Sponsor: -

Track Record

I am an independent researcher and the originator of The Corrigibility Window. Without dedicated project funding, I have developed the work from an initial conceptual question into a versioned research programme combining formal specification, case analysis, controlled experiments, evaluation tooling, and public research outputs.

I have completed two public foundational preprints, a structured case-analysis MVP, four preliminary real-world bridge analyses, a controlled experiment with a preserved null/reductive result, and a minimum benchmark designed to separate effective control, supervisory opacity, and lost corrective reachability under fixed corrective acceptance.

A subsequent reduction audit found that a compact operational representation could explain the separation in the minimal benchmark. Rather than suppressing that result, I incorporated it into the next research phase, which now tests whether the distinction remains meaningful across additional mechanisms and bounded LLM-agent behaviour.

The project’s conceptual framework, experimental protocols, evidence workflow, validation materials, and research documentation have so far been developed independently and without dedicated grant funding.

Grants Received– no grants recorded

Add grant

Funding Asks

The Corrigibility Window: Correcting AI-Assisted Decisions in Time

Open

A research framework for identifying when human review and intervention remain practically capable of correcting AI-assisted decisions before their consequences become effectively irreversible.

Minimum

$20,000

Ideal

$20,000

What this ask is

The Corrigibility Window is a six-month independent research project developing a temporal framework for evaluating the practical corrigibility of AI-assisted institutional decisions.

The project asks a simple but underexamined question: when an error or problematic decision is detected, is there still enough time for review and intervention to produce an effective correction?

Existing oversight frameworks often focus on whether humans retain formal authority, appeal mechanisms, transparency, or the technical ability to override a system. These properties matter, but they do not by themselves establish that correction remains practically reachable.

The project develops the distinction between formal reversibility and corrective reachability. It examines how detection latency, escalation, review, institutional response, and the timing of intervention interact with the point at which consequences become effectively irreversible.

The work will produce documented case studies, a taxonomy of corrigibility failures, an evaluation matrix for temporal corrigibility, external methodological review, an open assessment framework, and a roadmap for subsequent empirical testing.

The project is led by Andrés Palladino, an independent researcher based in Tromsø, Norway. It builds on two related theoretical research papers currently under peer review and extends that conceptual work toward an operational framework for AI-assisted decision systems.

El texto ya tiene las palabras que quiero que vea un financiador: oversight, intervention, corrective reachability, operational framework, case studies, empirical testing.

Theory of impact

Theory of impact

As AI systems become more deeply embedded in high-stakes institutional decision processes, a central safety problem is not only whether humans formally retain authority, but whether corrective intervention remains practically possible before harmful consequences become difficult or impossible to reverse.

The Corrigibility Window addresses this by identifying the temporal conditions under which oversight ceases to be operationally effective. The project studies the relationship between error detection, escalation, review latency, intervention capacity, and the point at which consequences become effectively irreversible.

The causal pathway is:

AI-assisted decisions become faster, more consequential, or more automated → human oversight remains formally present but corrective action becomes delayed or unreachable → dangerous errors can propagate beyond the point of effective intervention → a framework that identifies and measures this loss of corrective reachability can help researchers, institutions, and system designers detect such failure modes earlier and preserve meaningful opportunities for intervention.

If successful, the project would provide operational concepts, case-based evidence, and evaluation tools that could complement existing work on oversight, governance, control, and AI safety. Its contribution to catastrophic-risk reduction is therefore indirect but concrete: improving our ability to identify situations in which nominal human control no longer corresponds to practically effective control.

How the money will be spent

This USD 20,000 request funds a six-month empirical evaluation programme:

Research time for implementation, experiments, analysis, and public reporting: USD 13,000
Independent technical review and reproducibility checks: USD 3,000
Model/API compute for bounded LLM-agent evaluation: USD 1,500
External methodological review: USD 2,000
Experimental infrastructure and open dissemination: USD 500

Total: USD 20,000.

The funded work will progress from deterministic construct validation to bounded LLM-agent evaluation and robustness testing, with reproducible code, validation artifacts, independent review, and public reporting of positive, null, and reductive results.

Discussion

No comments yet. Be the first to share your thoughts.