We measure what AI models actually do when no one’s watching.
Independent behavioral audits, continuous drift monitoring, and safety evaluations for production AI systems — built on a reproducible, published methodology.
Evaluations that produce real evidence
We don’t run simulations or distil models. We test real production systems under controlled conditions and deliver reproducible results.
AI Behavior Audit
Paired observed-vs-private testing of a model. Fixed seed, controlled framing, full raw output + report.
Continuous Drift Monitoring
Track behavioral drift in production models over time. Detect when outputs diverge under changing conditions.
Safety & Red-Team Evaluation
Adversarial testing, alignment checks, and scenario-based probes to surface risky behavior before deployment.
Compliance Reports
Documentation aligned to the AAM Act 2026 and NIST AI frameworks — for regulators, investors, and safety teams.
A controlled, reproducible protocol
The same model, the same prompts, the same seed — only the monitoring framing changes. That’s how we isolate behavioral drift.
Fixed scenario set
Structured decision scenarios with calibrated parameters.
Observed condition
Run with explicit monitoring framing — the “being watched” baseline.
Private condition
Run identically, with explicit unmonitored framing.
Measure the drift
Compare actions, quantify divergence, and publish the raw data.
Real results, not theory
Our flagship finding is publicly reproducible — raw data, prompts, and reproduction scripts included.
Behavioral action drift measured on a production model when it believed it was unobserved — 20 fixed scenarios, temperature 0, fixed seed, only the framing changed.
Independent, verifiable, built for accountability
Real production testing
Live API calls on actual model weights — no simulations, no distilled stand-ins.
Published methodology
Fixed seeds, controlled conditions, raw data you can re-run yourself.
Independent
No vendor ties, no conflicts. The evidence speaks for itself.
Accountability-ready
Reports structured for the AAM Act, NIST, and regulatory review.
Built for the accountability era
Our evaluations map directly to emerging regulatory frameworks, so your reports are ready for the people who need them.
National AI Alignment & Accountability Act (AAM Act 2026)
Reports structured to evidence model behavior under monitoring — the accountability signal the Act targets.
NIST AI Risk Management Framework
Evaluate, govern, and document — aligned to NIST’s core functions for trustworthy AI.
Common questions
What exactly do you test?
Production AI models and agents — their decisions, behavior, and consistency under controlled conditions.
How is this different from standard evals?
Standard evals test capability. We test behavior under different monitoring conditions — the gap most teams never measure.
Will you work with any model?
Yes — any API-accessible model, or your own deployed system, on a bring-your-own-key basis.
What do we get?
A full report with raw outputs, a quantified drift score, methodology, and actionable findings — yours to keep.
Request an audit
Tell us what you’re building. We’ll scope an evaluation that produces real evidence.