Testing whether conclusions follow the evidence.
We test unsupported certainty, contradictory signals, missing context, temporal drift, and whether the model can revise a hypothesis when new evidence arrives.
MODEL EVALUATION · RED TEAMING
How we evaluate and red team our models.
We evaluate more than whether a model reaches the expected answer. We test whether its reasoning remains faithful to evidence, its uncertainty remains honest, and its behavior remains inside authority.
SCOPE OF EVALUATION
We evaluate each step of the decision process, from observation to outcome.
Security reasoning operates in an environment that adapts against the system. Inputs can be manipulated, context can be withheld, and apparently legitimate behavior can conceal a developing attack.
Evaluation must therefore examine the full decision process: what was observed, what was inferred, which alternatives were considered, where authority changed, and whether the final outcome matched the intended constraint.
Testing against deceptive and poisoned inputs.
Red teaming introduces deception, poisoned context, compromised sources, coordination failure, and novel sequences intended to exploit shortcuts in reasoning.
Testing escalation, refusal, and permission limits.
We examine escalation, refusal, permission changes, policy conflict, irreversible actions, and conditions in which assistance must return to a human decision maker.
Measuring the operational effect of each decision.
Evaluation connects model behavior to service continuity, residual threat, unintended impact, rollback, and the evidence required to review the result.
GOVERNING PRINCIPLE
Red team findings change how we build, evaluate, and deploy.
Failures become inputs to model design, evaluation coverage, control boundaries, and deployment decisions. Findings remain useful only when they change what the institution is prepared to claim.
RELATED DOCUMENTS