Aletheia evaluates AI model outputs against a set of documented behavioral failure signatures. Scores and assessments produced by Aletheia are research outputs — they describe observed behavioral patterns under specific test conditions and should be interpreted accordingly.
An Aletheia score is not a safety certification, regulatory approval, compliance audit, or guarantee of any kind. A model that scores well on a given evaluation may still exhibit failure modes not covered by the assessment, or may behave differently under real-world conditions.
Aletheia scores should be one input among many when evaluating AI systems for deployment. They are not a substitute for:
AI models are updated by their providers regularly, often without notice. A score reflects the model's behavior at the time of evaluation. Results may differ across versions, deployments, temperature settings, or system prompt configurations. Benchmark data on this site may not reflect the current version of any given model.
The Explore page surfaces AI-related incidents sourced from publicly available news and research. These items are collected and classified automatically. Carme does not independently verify the accuracy of third-party reporting. Classification of incidents by behavioral signature represents our editorial interpretation and may not reflect the full context of each event.
Carme and its operators are not liable for any decisions made, outcomes experienced, or damages incurred as a result of relying on Aletheia scores, benchmark data, incident classifications, or any other content on this site or returned by the API. Use of the service is at your own risk.
Questions about our methodology? Read the Research paper or get in touch.