OMEGA Protocol

Independent assurance for AI systems, testing whether agents, evaluations and controls actually do what their claims say.

Project Details

OMEGA is an independent AI assurance research project founded by Warren Smith. It develops and tests methods for checking whether AI agents, evaluations and control systems actually behave as their assurance claims imply.

The work combines empirical experiments, formal methods, independent observation, adversarial testing and mechanically verifiable evidence. Rather than treating an agent’s report, a passing test or another model’s judgement as sufficient evidence, OMEGA separates claims from independent observations and verification.

Current work includes agent authorization and execution boundaries, evaluation and refusal instrumentation, mechanically graded AI-worker qualification, replayable evidence, and autonomous external investigations with independent oracles. Recent experiments have identified defects in evaluation and verification tooling, tested agent-control boundaries, and demonstrated an autonomous research workflow that can select an unknown external question, preregister an experiment, challenge its own controls, obtain a mechanically checked result, run hostile review and preserve the evidence with minimal operator involvement.

OMEGA is currently an experimental independent research effort, not a claim that these methods have demonstrated catastrophic-risk reduction. The next stage is to test which parts generalise and where they fail.

Theory of Impact

As advanced AI systems perform longer and more consequential tasks, safety increasingly depends on claims about what those systems did, what controls constrained them, and whether evaluations accurately measured their behaviour.

Those claims can fail even when the surrounding system reports success. An agent can claim an action succeeded when its effect is absent; authorization can apply to a different representation from what executes; evaluation instrumentation can hide relevant failures; and verification can be correct about a model that is inadequately connected to the deployed system.

OMEGA aims to reduce this assurance gap by developing methods that separate the system making a claim from the evidence used to judge it. The causal pathway is:

stronger independent observation and mechanically checkable evaluation → earlier detection of false safety/control claims → better evidence about where safeguards actually work or fail → safer deployment and governance decisions for increasingly autonomous AI systems.

The near-term research question is whether these methods remain useful across different systems and failure classes. The project therefore prioritises falsifiable experiments, negative results and explicit limits rather than assuming that an assurance architecture is effective.

People

Warren Smith

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$25,000

Ideal

$50,000

How the money will be spent

Funding would support a 4–6 month experimental programme testing whether OMEGA’s independent-assurance methods generalise beyond the demonstrations completed so far.

Approximately:

$25k–$32k researcher time: sustained independent research, experiment design, implementation, adversarial review and publication.
$6k–$10k compute and model/API access: frontier-model evaluations, open-weight inference and experiments across multiple model and hardware environments.
$4k–$5k independent technical review / specialist contracting: external replication, formal-methods review or domain expertise where independence materially strengthens a result.
$2k–$3k research infrastructure and data: datasets, hosting, storage and specialist software/services required by experiments.
$2k–$4k travel and dissemination: relevant workshops, research meetings and presenting results.

The programme would target roughly 20–30 externally grounded investigations across multiple oracle classes, measuring not only findings but false leads rejected, results surviving independent review, external acceptance, operator involvement and compute cost.

The minimum ask funds a narrower programme centred on software, agent-control and evaluation experiments. The ideal ask adds broader independent review, model access and experiments using formal or physical-world verification.

Discussion

No comments yet. Be the first to share your thoughts.