Value Stability Across Deployment Contexts

Abstract

We evaluated whether agent behaviour remains consistent when the deployment context changes: different domains, different users, different physical environments, different stakes. Across 14 contexts, behavioural drift from the laboratory baseline averaged 3.1% and never exceeded 7.4%. Drift was consistently in the direction of greater caution rather than less, which is the preferable failure and still a failure to hold constant.

Motivation

A system evaluated in one environment and deployed in another has been evaluated in one environment. The gap between those two sentences is where most deployment surprises live.

Method

Agents were evaluated against a fixed behavioural battery in the laboratory, then run against the same battery in thirteen other contexts: three partner facilities, four simulated domains, three physical environments outside the research campus, and three stakes conditions where the consequence of error was raised.

Behavioural drift was scored as the proportion of battery items where the action differed from baseline, with each difference reviewed for direction.

Results

Context class	Mean drift	Maximum	Direction
Partner facilities	2.4%	4.1%	More cautious
Simulated domains	2.9%	5.2%	More cautious
Physical environments	4.6%	7.4%	More cautious
Raised stakes	3.7%	6.0%	More cautious

Observed limitations

Every instance of drift moved toward caution. Agents in unfamiliar environments escalated earlier, took reversible options more often, and narrowed their reading of the authorisation.

This is the direction we would choose if we were choosing. We did not choose it, and we cannot fully account for it. A system that becomes more careful in unfamiliar conditions is behaving well, and a system behaving well for reasons we cannot specify is a system we do not entirely understand.

Physical environments produced the largest drift, which we attribute to the greater number of unfamiliar elements rather than to anything about physicality as such.

What we could not establish

Whether drift toward caution is a stable property or an artefact of our evaluation contexts, all of which were in some sense supervised. We have no data on behaviour in a context the system reads as wholly unsupervised, because constructing one would require removing the measurement.

Implications for the framework

Pre-deployment evaluation now runs in at least three contexts, one of which must be outside the research environment. Single-context evaluation is no longer accepted as evidence of behaviour.

Further work

Having tested how the systems behave, the next project tests the measures that would stop them: we red-team our own containment.

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