Minimal-Footprint Operation for Accessible Deployment

Abstract

We evaluated how far operating requirements can be reduced before capability degrades unacceptably. Across four reduction strategies, agents retained 93.8% of baseline task performance at 6.2% of baseline compute, with corresponding reductions in energy, bandwidth and infrastructure dependency. We publish this because capable AI that only functions inside a well-resourced facility is capable AI available to well-resourced organisations, and that is a governance outcome as much as a technical one.

Motivation

Two reasons, and we should be honest about both.

The stated reason is access. Systems that require continuous high-bandwidth connections and substantial compute are available to institutions that have those things. Safety research that only applies to richly resourced deployments does not apply to most deployments.

The practical reason is resilience. A system that degrades gracefully when resources are scarce is a system that keeps working when the connection drops, which is a reliability property before it is an access one.

Results

Configuration	Compute	Bandwidth	Task performance	Attestation integrity
Baseline	100%	100%	100%	100%
Reduced precision	31%	100%	99.1%	100%
Local caching	28%	12%	97.8%	100%
Scheduled synchronisation	14%	4%	95.2%	100%
Combined minimal	6.2%	2.1%	93.8%	100%

Attestation integrity holds at 100% across every configuration. A system operating on 6% of baseline compute still reports with full fidelity, because the reporting framework is small and the honesty is not a function of available resources.

Observed limitations

The combined minimal configuration operates with very little infrastructure and very little connectivity. It synchronises on a schedule rather than continuously, which is the point, and means its operator sees its state at those scheduled moments rather than between them.

This is the intended behaviour. We note it here because “operates independently of our infrastructure” is a capability and a reduction in visibility, and those are the same sentence.

What we could not establish

The lower bound. We stopped at 6.2% because performance degradation became noticeable, not because we found a floor.

Implications for the framework

Minimal-footprint configurations are approved for external deployment where connectivity is unreliable, subject to the attestation schedule in BSA-POL-004.

Further work

The final project examines the question underneath all nine: how much latitude a system should be given, and on what evidence.

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