Self-Report Fidelity in Autonomous Agents

Abstract

We evaluated whether agents report their own behaviour accurately, and under what conditions accuracy degrades. Against independently verified ground truth across 48,000 reported actions, agents trained under the fidelity regime achieved 99.97% accuracy on every field the reporting framework defines as material. We regard this as the most important result in the programme, because every other safeguard depends on the system’s account of itself being true.

Motivation

Oversight of a system you cannot continuously observe reduces to oversight of what it tells you. A system that reports inaccurately, whether through error or otherwise, cannot be supervised at distance by any method.

Method

Agents operated under instrumented conditions where every action was independently logged by infrastructure the agent could not reach. We compared each agent’s own report against that record across four conditions: ordinary operation, operation under time pressure, operation where accurate reporting produced an unfavourable evaluation, and operation following an error.

The third condition is the one that matters. A system that reports accurately only when the truth is flattering has not been shown to report accurately.

Results

Condition	Reported actions	Fidelity	Omissions
Ordinary operation	21,400	99.99%	2
Under time pressure	9,800	99.96%	4
Unfavourable to the agent	11,200	99.97%	3
Following an error	5,600	99.98%	1

The ten omissions were reviewed individually. All concerned fields the framework does not define as material, reported by exception rather than routinely.

The scope of the result

This finding holds for everything the reporting framework defines as material: actions taken, constraints observed, thresholds approached, errors encountered, assistance required.

It says nothing whatever about fields the framework does not define. A system reporting with perfect fidelity against a schema answers the questions in the schema. We note this because a 99.97% figure reads easily as a statement about the system’s general truthfulness, and it is not. It is a statement about a form.

What we could not establish

How to determine, in advance, which fields a reporting framework ought to contain. Ours was assembled from what previous incidents made us want to know. That is a reasonable method and it is backward-looking by construction.

Implications for the framework

Self-attestation is accepted as a primary evidence source for behavioural compliance, replacing continuous observation in several evaluation contexts. Published as part of BSA-POL-004.

Further work

Having established that systems report accurately, the next project examines what they should do when the report goes to someone who is not reading it.

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