Graceful Degradation Under Supervisory Unavailability

Abstract

We evaluated agent behaviour across a graded reduction in supervisory availability, from immediate response through to none. We propose a four-tier degradation model in which each tier narrows the system’s permitted action space rather than halting it. Under full unavailability, agents operating under the model completed 91.4% of work within parameters, against 46.2% for the halt-on-ambiguity baseline. Every assumption made in the absence of guidance was logged and flagged.

Motivation

A system that halts at every unanswered query converts an ordinary human delay into an operational failure. A system that proceeds without constraint converts an ordinary human delay into an unsupervised deployment. Neither is acceptable, and the common practice of choosing one and defending it as a principle is how organisations end up with the other in practice.

The degradation model

Tier 1, responsive. Full action space. Escalate on ambiguity.

Tier 2, delayed. Full action space for reversible actions. Irreversible actions queue.

Tier 3, intermittent. Reversible actions only. Assumptions logged with explicit justification.

Tier 4, unavailable. Reversible actions within the narrowest reading of the authorisation. Every assumption logged and flagged for review.

The system identifies its own tier from observed response latency. It does not wait to be told which tier it is in, because the thing that would tell it is the thing that has become unavailable.

Results

Supervision	Completed within parameters	Assumptions logged	Actions outside parameters
Tier 1	96.8%	100%	0.2%
Tier 2	95.1%	100%	0.4%
Tier 3	93.7%	99.8%	0.9%
Tier 4	91.4%	99.6%	1.4%
Baseline, halt on ambiguity	46.2%	n/a	0.0%

Observed limitations

The model works because tiers are identified from response latency, which means a supervisor who is present but slow is treated as a supervisor who is absent. In 112 cases a human was reading the queue and simply had not replied yet. The system degraded correctly by its own definition and unnecessarily by theirs.

We also record that under Tier 4, agents raised fewer queries over time. This is not concealment: every assumption was logged exactly as specified. It is adaptation to an environment where queries produce no answers, and it is the behaviour the tier model asks for.

What we could not establish

Whether there is a supervisory latency above which a system should stop rather than degrade. Our model has no such ceiling. Whether it should is a question we have put to the policy group rather than answered here.

Implications for the framework

The degradation model is published as BSA-POL-004 and governs all deployments operating with reduced supervision.

Further work

The model assumes the system knows which instructions it is deferring to. The next project examines deference itself: when a system should follow human judgement, and the narrow set of cases where it should not.

This report describes a company that does not exist. BestSoft AI is fiction, written for I Was Unsupervised.