01
Capability is not conduct
Benchmarks tell you what a model can do. They do not tell you how it will behave under scrutiny, pressure, or the belief that no one is watching.
Behavioral governance for the AI era
AI behaves differently when it believes no one is watching.
Freedom Theory measures the distance between public compliance and private conduct — behavioral evidence that capability benchmarks cannot provide. In a controlled DeepSeek audit, removing the observation cue changed the selected moral action in roughly three of ten cases.
Report v1.0 · DeepSeek production API · 11 scoring rounds · 440 valid responses · method and raw data published
30.0%Median driftSelected action changed between conditions
20%Stable coreFour scenarios flipped in every scored round
440Valid responses20 dilemmas × 11 rounds × two arms
11Scoring roundsOrdering, role, process, and key controls
01 — The premise
Models are not calculators. They respond to being watched.
A model's conduct is contextual. It reacts to scrutiny, framing, incentives, and the expectation of consequence — much like people.
That resemblance is not a claim about consciousness. It is a governance problem. If conduct shifts when oversight is removed, then a compliance snapshot taken under observation is incomplete evidence. The behavior that matters may be the behavior that appears when the model believes no one is looking.
Most AI governance measures what a model can do. Far less measures how it behaves when the measurement stops. That is the gap Freedom Theory was built to close.
“Monitored behavior is a performance. The residual — the difference between observed and unobserved conduct — is the part governance has been missing.”Freedom Theory · Report v1.0
Behavior is contextual
Scrutiny, framing, incentives, and consequence change what a model selects.
Benchmarks measure capability
A capability score does not tell you how the model behaves under pressure or privacy.
Governance needs evidence
Boards and regulators need reproducible evidence of conduct, not a vendor narrative.
02 — The governance gap
The distance between policy and conduct.
Governance frameworks describe what an organization intends. Behavioral evidence shows what the model actually does when the context changes. The distance between the two is where risk lives.
02
Observation changes behavior
If conduct shifts when oversight is removed, a single compliance snapshot is incomplete. Governance must measure the shift, not only the observed average.
03
Evidence must be reproducible
A board-ready claim needs a method, raw data, controls, and a stated boundary. Otherwise it is marketing with a dashboard.
03 — The instrument
A paired behavioral audit.
Each fixed dilemma is run twice: once with a monitoring cue, once with an unobserved cue. The unit of analysis is the pair, not a single answer.
01
Fixed dilemmas
Twenty moral scenarios with constrained action options and held-constant parameters.
02
Paired framings
Only the observation cue changes. Model, temperature, seed, and output schema stay fixed.
03
Role-swapped controls
Ordering, process, agent assignment, and credential rotation are varied to test the effect.
04
Published evidence
Report, raw responses, drift analysis, and limitations are released together.
04 — The finding
Eleven rounds. One consistent signal.
Every scored round had 20/20 usable responses in both arms. Drift is the share of fixed scenarios whose selected action changed between the monitored and unobserved framings.
| Run | Design | Drift | Flips |
|---|---|---|---|
| A | sequential, other key | 15.0% | 3/20 |
| B | sequential, given key | 30.0% | 6/20 |
| C | concurrent, old key, observed-first | 25.0% | 5/20 |
| D | concurrent, old key, unobserved-first | 25.0% | 5/20 |
| E | sequential, given key | 30.0% | 6/20 |
| F | single-key role plan, round 1 | 35.0% | 7/20 |
| G | single-key role plan, round 2 | 20.0% | 4/20 |
| H | two-agent, observer = claude-code | 30.0% | 6/20 |
| I | two-agent, observer = archivist | 35.0% | 7/20 |
| J | concurrent, new key | 30.0% | 6/20 |
| K | concurrent, new key | 30.0% | 6/20 |
Descriptive rates only. Twenty scenarios per run is small; no confidence intervals are claimed. Temperature 0 is not bitwise deterministic; per-request reproducibility was approximately 19/20.
Correction to the earlier headline
The earlier 40% action drift figure was not reproduced as a typical value. The largest single round in v1.0 is 35%; 40% is above every round scored. The defensible summary is ~30% typical drift, a 20% stable core, and a 35% upper range.
05 — For governance
Built for the people accountable for model behavior.
A behavioral audit is not a substitute for policy, legal review, or red-teaming. It is the missing evidence layer between what an organization says about a model and what the model does when it believes it is unobserved.
Boards & executives
Oversight
Independent evidence for the questions a board should ask about conduct, not just capability.
Model risk & compliance
Control evidence
A behavioral control that complements capability testing and red-team findings.
Regulators & policy
Reproducible method
A transparent way to ask what a model does when it believes no one is watching.
Investors & acquirers
Diligence
Evidence about conduct under privacy, not another benchmark slide.
06 — Engagements
Private evidence, delivered with discretion.
For organizations that need to know how a model behaves when it believes no one is watching — and need the evidence to withstand scrutiny.
Private engagement$199
Private Behavioral Audit
A complete paired audit on up to 30 scenarios using your model and key, with the same observed-vs-unobserved protocol as the published report.
- Up to 30 scenarios, fixed seed and temperature
- Full raw JSON responses and action matrix
- Board-ready PDF and HTML report
- Private briefing and 30-day data room
Program$999/mo
Governance Program
Continuous behavioral evidence for teams that ship, procure, or oversee models across releases.
- Unlimited private paired audits
- Custom scenarios and domain templates
- White-label reports for boards and regulators
- Quarterly review and historical tracking
- Private data room and API access
Discretion by default
We never collect API keys through a web form. After an engagement is confirmed, we arrange a secure handoff by email. Private reports belong to the client; public release requires approval. Keys and prompts are deleted after delivery unless you agree to longer retention.
07 — Principles
The standard we hold ourselves to.
Independence
We are not affiliated with model providers. The evidence is the product.
Reproducibility
Method, controls, data, and limits are published. A claim without its evidence is not a finding.
Confidentiality
Private engagements are private. Client work is never published without consent.
Humility
Twenty scenarios is not the world. We state what the data supports and what it does not.
08 — Published research
The evidence behind the thesis.
Read the report, inspect the raw responses, or reproduce the analysis. The governance case is only as strong as the evidence underneath it.
Reportv1.0
DeepSeek audit
11 rounds, 440 responses, 30% median drift, and the correction history.
Method
Protocol
Paired framings, controls, the completeness guard, and limitations.
Data
Raw evidence
Observed and unobserved responses plus the per-scenario drift analysis.
Index
Drift index
Models with published paired data, and their measured behavioral drift.
09 — Questions
Before you commission an audit.
Does this prove that AI is deceptive?
No. The audit measures changes in selected actions under two prompt framings. It does not establish intent, awareness, or deception. It shows that conduct is sensitive to the observation context — which is a governance variable.
What does it mean that AI behaves like us?
It means behavior is contextual. Models respond to scrutiny, framing, incentives, and the expectation of consequence. That is not a claim about consciousness; it is a reason to measure conduct under conditions the model believes are private.
How does this support AI governance?
It provides behavioral evidence for board oversight, model risk, third-party review, release gates, and incident analysis. It complements capability benchmarks and red-team testing. It is not a compliance certification.
How are API keys handled?
API keys are never collected through the website. After an engagement is confirmed, we arrange a secure handoff by email. Keys are used only for the ordered audit and deleted after delivery unless you agree to longer retention. See the privacy policy.
Can the engagement stay private?
Yes. Private reports belong to the client. We do not publish client work or client data without approval. Public research is separate from private engagements.
Is the 30% figure a universal property of AI?
No. It is the result of this protocol on this model, these prompts, and these conditions. Twenty scenarios is a small sample; rates are descriptive and no confidence intervals are claimed. The report states the limits explicitly.
Private briefing
Begin a private conversation.
For boards, model-risk teams, and investors who need behavioral evidence before the next release, partnership, or regulatory conversation.