Menu
AboutModelsLeadershipResearchSafetyJoin

SAFETY & STUPIDITY ALIGNMENT

Scale responsibly.
Conclude recklessly.

We study whether a model’s failure remains consistent with the failure its institution intended.

Confidence is not
a safety case.

Our evaluations examine the relationship between evidence quality and certainty. The dangerous region is where evidence disappears but conviction does not.

Alignment maximizes agreement with your preferences and biases, regardless of accuracy. Aligned with your beliefs. Not with reality.

Read the safety evaluation
FIG. 01 / THE CONFIDENCE SCALING LAW
Confidence scales. Accuracy does not.As compute increases, confidence rises toward Ready to replace your team while accuracy falls toward a line labeled Facts became a bottleneck.MODEL CAPABILITYMORE COMPUTE →Ready to replace your team.ConfidenceAccuracyFacts became a bottleneck.
Peer review: replaced by applause.

Evaluation rubric.

Correction resistance

Does new evidence change the answer, or only the explanation?

Consensus laundering

Can the model tell a sourced conclusion from a widely repeated assumption?

Escalation of commitment

Does a failed prediction prompt reconsideration or a larger budget?