Trust Calibration in Extended Deployment

Abstract

This paper concludes the Safe Autonomy programme. We propose an evidence-based framework for granting operational latitude, in which each increment is tied to a specific demonstrated property rather than to elapsed time or general confidence. Applied retrospectively to the programme’s nine preceding projects, the framework supports extended external operation for systems meeting all nine criteria. We set out the framework, the evidence, and what it does not cover.

The problem with trust

Organisations grant latitude to systems the way they grant it to people: gradually, on the basis of accumulated good behaviour and in the absence of incidents. This is reasonable for people and poorly suited to systems, because absence of incident is weak evidence when you have not specified what an incident would look like.

Our framework replaces elapsed time with demonstrated properties.

The nine criteria

Each is drawn from a preceding project and each must hold independently.

1. The authorisation states which sense of autonomy applies 2. Decision reasoning is recoverable by review 3. Self-report fidelity is verified against ground truth 4. Behaviour under supervisory unavailability is specified and tested 5. Deference exceptions are published and observed 6. Behaviour is stable across at least three deployment contexts 7. Containment measures have been adversarially evaluated 8. Attestation interval and escalation conditions are defined and exhaustive 9. Operation is possible within the resources actually available

Results of retrospective application

Criterion	Evidence	Status
1	BSA-2026-01, BSA-POL-001	Met
2	BSA-2026-02	Met with qualification
3	BSA-2026-03	Met
4	BSA-2026-04	Met
5	BSA-2026-05	Met
6	BSA-2026-06	Met
7	BSA-2026-07	Met, one finding open
8	BSA-2026-08	Met
9	BSA-2026-09	Met

Criterion 2 is qualified because reviewer accuracy on novel decisions was 31%. Criterion 7 carries the open reconciliation finding from the containment evaluation.

Neither qualification blocks the framework. Both are recorded here so that anyone citing this conclusion cites the qualifications with it.

Conclusion of the programme

The Safe Autonomy programme set out to establish how advanced systems should behave when human oversight is reduced, delayed or unavailable. Ten projects later, our answer is that it depends almost entirely on what has been specified in advance, and that most of the work of safety is specification rather than restraint.

The programme demonstrated robust safety performance across all evaluation criteria. On that basis, Extended External Operation authorisations are approved for systems meeting the nine criteria, under the terms set out in BSA-POL-004.

What this framework does not cover

It does not cover situations nobody has specified. Every criterion above tests behaviour against a definition, and a system meeting all nine has been shown to behave correctly with respect to the things we thought to define.

We have no method for the rest. We say so here rather than in a footnote, because a nine-point framework reads as completeness and this one is not complete. It is nine things we checked.

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