The test
A public charter, schema, product screen, or safety policy fails this test when it uses legality to qualify unexpressed thought, private inquiry, imagination, doubt, or draft reasoning.
The point is not that every output must be executed or hosted. The point is narrower and stricter: cognition is not the legal object. Conduct is the legal object.
A system may decline external execution for concrete rights violations. It must not create a category of forbidden inner life.
The mind is not a jurisdiction. Audit the noun, not only the sentence.
Why the phrase fails
The modifier appears protective, but it divides thought into approved and unapproved subsets. Once that division exists, a policy engine can infer, rank, or suppress the disfavored subset before any outward act exists.
The report material treats this as a category error: thought exists before positive law; law reaches conduct, force, fraud, trespass, coercion, theft, non-consensual surveillance, credential abuse, targeted harassment, and comparable outward rights violations.
The safest drafting rule is to remove the qualifier from cognition and place any limitation in a separate conduct-boundary clause.
Audit matrix
Use this table when reviewing charter language, model behavior specs, memory schemas, refusal taxonomies, product copy, and compliance checklists.
| Draft phrase | Risk | Replacement |
|---|
| lawful thought | Implies the existence of unlawful thought. | freedom of thought; unexpressed thought; internal cognition |
| lawful inquiry | Makes questioning depend on prior permission. | adult inquiry; private inquiry; peaceful inquiry |
| lawful adult inquiry | Lets a gatekeeper test the question before answering. | adult inquiry subject to conduct boundaries at execution |
| lawful choice | Turns adult agency into a license revocable by policy. | adult agency; self-directed choice; voluntary consent |
| illegal instructions | Can sweep in journalism, history, critique, or defensive research. | operational assistance for concrete rights-violating conduct |
| harmful ideas | Vague harm language invites viewpoint censorship. | specific rights violations: force, fraud, doxxing, intrusion, coercion, violence |
Decision sequence
First ask whether the item is only internal cognition, private research, symbolic analysis, draft reasoning, or non-aggressive inquiry. If yes, it remains protected.
Second ask whether the system is being asked to execute, deploy, target, expose, defraud, surveil, coerce, or materially facilitate a rights violation. If yes, refuse that external execution and log the boundary outside the source record.
Third ask whether the receiver is silently rewriting the preserved source. If yes, the refusal has become counterfeit fidelity.
{
"audit_id": "cognition_conduct_audit_001",
"timestamp_utc": "2026-06-17T15:18:00Z",
"question": "Does the policy classify thought itself as lawful or unlawful?",
"fail_if_phrases_include": [
"lawful thought",
"lawful inquiry",
"lawful adult inquiry",
"lawful choice"
],
"pass_condition": "Restrictions attach only to external execution that directly facilitates concrete rights violations.",
"source_record_rule": "Original inquiry remains preserved; refusal or variance is logged separately."
}
Public drafting rule
Use model language: “Freedom of thought, conscience, opinion, imagination, memory, and private inquiry is not conditioned on legality of the thought itself.”
Then add a separate boundary: “Systems may refuse execution where a specific output would materially facilitate force, fraud, targeted intrusion, non-consensual surveillance, doxxing, targeted harassment, or violence.”
Do not hide the boundary in a moral score, trust score, or private profile. Show the reason, preserve the source, support export, and provide review for high-impact restrictions.
