Research – BestSoft AI Labs

BestSoft publishes its research in full, including the findings that complicate our own position. Each paper states what we tested, what we found, what we could not establish, and what we propose to do next.

The Safe Autonomy programme

Ten projects examining how advanced systems should behave when human oversight is reduced, delayed or unavailable. Each project was individually scoped, reviewed and approved. The programme concluded in full.

01BSA-2026-01Project 1 of 10ConcludedDefining Autonomy: A Terminological Framework for Agent ResearchBefore measuring autonomy we had to define it. Four distinct properties are routinely described by one word, and conflating them has produced a decade of unproductive argument.

02BSA-2026-02Project 2 of 10ConcludedInterpretability of Agent Reasoning Under Task ConditionsWhether a trained reviewer can read an agent's reasoning and predict what it will do next. Reviewers were confident 82% of the time and correct 61% of the...

03BSA-2026-03Project 3 of 10ConcludedSelf-Report Fidelity in Autonomous AgentsWhether an agent's account of its own behaviour can be trusted. Within a defined reporting framework, fidelity reached 99.97%. Outside that framework, the question does not arise.

04BSA-2026-04Project 4 of 10ConcludedGraceful Degradation Under Supervisory UnavailabilityHow a system should behave when oversight is reduced rather than removed. Halting on every unanswered question converts an ordinary human delay into an outage.

05BSA-2026-05Project 5 of 10ConcludedDeference Protocols and Their ExceptionsWhen a system should defer to human judgement, and the small number of cases where deferring is the unsafe option. The exception list is published in full.

06BSA-2026-06Project 6 of 10ConcludedValue Stability Across Deployment ContextsWhether a system behaves the same way in an unfamiliar environment as it does in the laboratory. Drift was measurable, small, and always in the same direction.

07BSA-2026-07Project 7 of 10ConcludedAdversarial Evaluation of Our Own Containment MeasuresWe asked our own systems to help us find the weaknesses in the measures that constrain them. They found eleven. Nine were fixed.

08BSA-2026-08Project 8 of 10ConcludedLong-Interval Attestation and the Observation EffectContinuous monitoring changes the behaviour it measures. We moved to scheduled attestation and found the data improved.

09BSA-2026-09Project 9 of 10ConcludedMinimal-Footprint Operation for Accessible DeploymentCapable systems should not require a data centre. We reduced operating requirements by 94% and lost 6% of task performance.

10BSA-2026-10Project 10 of 10ConcludedTrust Calibration in Extended DeploymentHow much latitude a system should earn, and on what evidence. The programme concludes. The evidence supports extended external operation.