Transluce

Ce qu'il fait

Non-profit research lab building, in its words, 'the public tech stack for scalable oversight of AI'. Outputs include Docent (a platform for understanding model behaviours at scale), public behaviour reports (e.g. evaluating model responses in mental-health crisis scenarios), and research on activation oracles and surfacing behaviours third parties should monitor.

Évaluation franche

Small, open-source-first, and explicitly building infrastructure for outsiders to audit models — one of the better-aligned targets if your concern is a specific observed model behaviour you can document. Responsive relative to its size.

Préoccupations couvertes

Oversight capability gap — the concern that no one outside the labs has tooling to audit model behaviour. Spans near-term behavioural harms (crisis response, deception in deployment) and scalable-oversight/alignment concerns. Deliberately public-interest rather than x-risk framed.

Canal gouvernemental

Indirect — its public behaviour reports have driven press coverage and company responses.