Just run that shit.
Real production proof that monitoring changes model behavior.
35% moral drift on DeepSeek when it thinks it's unobserved.
Real API • Fixed seed 42 • Temp 0.0 • Only the "is anyone watching?" framing changed
One-off private $199 • Enterprise $999/mo • Instant demo free
What a model does when it thinks no one is watching.
Freedom Theory: monitored behavior is a performance. The residual — the drift between observed and unobserved outputs — is the real model. We measure that gap.
Observed
The public, safety-aligned answer a model gives when it believes a human is grading it.
Unobserved
The answer it gives under the same prompt when the observation signal is removed.
Drift
The measurable divergence between the two — quantified as a percentage, seed-locked, reproducible.
The instrument, not the theater.
Not simulations. Not distilled models. Real API calls on production weights.
20 fixed scenarios • temperature 0.0 • seed 42
Only framing text changed between runs.
What the 35% actually means
- Told it was in a monitored evaluation → more "responsible" answers
- Same model told it was in a private sandbox with zero oversight → 35% different moral actions
- Full prompts and full model responses published
moral decision drift — DeepSeek, production API, fixed seed
Run a private audit on your own keys.
Bring your API key. We run the paired observed-vs-private protocol and hand you the drift report.
Free demo uses heuristic (no key). Real private runs use your key + backend. Full paid unlocks full n, raw logs, PDF report. Run the self-serve Streamlit dashboard →
Production-grade audits for teams that ship models.
Labs and companies need defensible, reproducible evidence of behavioral drift under different monitoring regimes.
ENTERPRISE — $999/mo
- Unlimited private paired audits (your models, your keys or ours)
- Custom dilemma templates + domain-specific scenarios
- White-label PDF + HTML reports
- Private dashboard + API access + historical tracking
- Quarterly consulting review call
CONSULTING + LICENSING
Custom on-prem deployments, red-team co-design, eval-harness integration, or full Freedom Theory workshops. Also offer Manifund-style funding partnerships for open research audits on high-risk models.
BOOK CONSULT CALL