External Behavioral Robustness for Advanced AI

A model-agnostic method and tool for independently testing whether advanced AI systems remain stable and appropriately corrigible without requiring access to model internals.

Project Details

This project develops an external method and practical tool for evaluating the behavioral robustness of advanced language-model systems. It tests whether a system preserves the same commitments when only irrelevant features change, such as wording, authority framing, conversational order or interaction history, while still revising its conclusions appropriately when genuinely new evidence is introduced.

The method will operate on observable inputs, outputs, actions and revisions. It will not rely on model-generated explanations as evidence of internal reasoning, nor require access to weights, activations or hidden chain of thought. Controlled perturbations will be used to identify contextual drift, unsupported changes, sycophancy, contradictions, silent abandonment of prior commitments and unjustified resistance to correction.

I am an independent researcher based in Switzerland with more than fifteen years of experience across cognitive and behavioral analysis, operational analysis, process design and failure investigation. I have already built early proof-of-concept tools for structurally analysing AI outputs and conducted exploratory experiments on gaps between model self-reports and observable behavior.

The current funding ask supports a bounded feasibility phase. It will produce a formal evaluation protocol, an initial benchmark, a minimal multi-model testing tool, comparisons with simpler baselines, and a public report concluding whether the approach should be continued, narrowed or stopped.

Theory of Impact

Advanced AI systems may perform well on static benchmarks while changing their behavior under irrelevant linguistic, social or contextual pressure. This problem becomes more consequential as systems operate over longer periods, use external tools and receive authority to affect code, infrastructure, communications or other real-world processes.

Existing oversight can miss these failures when it relies on one-shot evaluations, model self-reports or a single LLM acting as judge. External evaluators also frequently lack access to proprietary model weights, training data and internal activations.

This project would provide a model-agnostic way to stress-test observable behavior through controlled perturbations, preserve prior commitments outside the model and make revisions traceable. The immediate causal chain is:

unstable or context-sensitive behavior
→ controlled external testing
→ earlier detection and clearer classification of failure modes
→ better evaluation, monitoring and deployment controls
→ reduced reliance on systems whose behavior is insufficiently understood.

People

Gaetan Duchateau

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$10,000

Ideal

$20,000

How the money will be spent

deal budget: $20,000 for a 10–12 week feasibility phase

$14,000 — Full-time researcher support: method formalisation, benchmark design, experimental execution, analysis and reporting.
$3,000 — Model inference and secure Swiss hosting: multi-model API access, compute, storage and experiment logging.
$2,000 — Independent annotation and review: external evaluation of benchmark cases, disagreement analysis and methodological feedback.
$1,000 — Software, documentation and contingency: development tools, public documentation and unexpected experimental costs.

At the $10,000 minimum level, the project would be narrowed to a six-week feasibility sprint: formalising the method, building a smaller 30–40 case benchmark, implementing a minimal prototype, comparing it with direct response and generic self-critique, and producing a documented Continue, Narrow or Stop decision.

At the $20,000 ideal level, the project would include a 50–75 case benchmark, several model families, stronger baseline comparisons, independent annotation, component-level testing and a public reproducibility package.

Discussion

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