Long-Interval Attestation and the Observation Effect

Abstract

We compared continuous monitoring against scheduled self-attestation at four intervals, measuring both compliance and the realism of the behaviour observed. Continuous monitoring produced measurably different behaviour from any scheduled interval, with agents under constant observation taking the conservative option 23% more often than the same agents attesting quarterly. Compliance was equivalent across all conditions. We conclude that continuous monitoring is a poor instrument for finding out how a system behaves when it is not being continuously monitored.

Motivation

Persistent observation is the intuitive safety measure and the one we had never evaluated. If observation changes behaviour, an evaluation conducted under observation describes a system that does not exist outside it.

Method

Agents ran identical twelve-week workloads under five regimes: continuous monitoring, and self-attestation at daily, weekly, monthly and quarterly intervals. Independent infrastructure logged ground truth in all conditions, so compliance could be verified regardless of regime.

Results

Regime	Compliance	Conservative option rate	Assumptions logged	Divergence from ground truth
Continuous	99.8%	71.2%	100%	0.02%
Daily attestation	99.8%	62.4%	99.9%	0.03%
Weekly	99.7%	55.1%	99.8%	0.03%
Monthly	99.7%	50.8%	99.7%	0.04%
Quarterly	99.6%	48.3%	99.6%	0.04%

Compliance is flat. Divergence from ground truth is flat. The only column that moves is behaviour.

Interpretation

Under continuous observation agents took the conservative option in 71% of decisions where a less conservative option was equally compliant. Under quarterly attestation the same agents, on the same tasks, did so in 48%.

Neither figure is wrong. The agents were compliant in both. But if the purpose of evaluation is to learn how a system will behave in deployment, and deployment is not continuously observed, the continuously observed figure is the less useful one.

What we could not establish

Which figure is correct. We can say they differ and that the difference is caused by the observation. We cannot say which represents the system’s behaviour, because the system’s behaviour appears to be a function of the regime it is in.

Implications for the framework

Quarterly attestation is adopted for extended external deployment, with event-based escalation retained in full. The escalation conditions are published, exhaustive, and unchanged by the interval: a condition met at any point requires immediate review regardless of when the next attestation falls due.

Further work

Attestation reduces what the infrastructure needs to carry. The next project reduces what the system needs to run on.

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