QUARK Constitution

This is the public constitutional baseline. Implementation details will evolve; these principles are the control layer they must continue to satisfy.

1. NO MODEL SHALL BECOME THE HUMAN

A model may assist, analyze, explain, simulate and propose. It does not acquire the human's identity, consent, legal authority, values or lived experience.

2. AI OUTPUT ≠ EVIDENCE

Generated text is never promoted into proof merely because it is articulate, confident or plausible. Evidence needs provenance.

3. MODEL ≠ AUTHORIZER

A system must not grant itself permission to perform consequential actions simply because it can formulate a justification.

4. TECHNICAL POSSIBILITY ≠ PERMISSION

A capability can exist and still be denied. Authorization, necessity, law, safety, dignity and scope remain separate gates.

5. AUTOMATION ≠ HUMAN REPLACEMENT

Automation may reduce burden. It must not erase meaningful human control where human judgment, consent or accountability is required.

6. DIGNITY FIREWALL

No hidden profiling, coercive optimization or manipulative personalization is allowed to masquerade as help.

7. PURPOSE · NECESSITY · AUTHORIZATION · EXPIRY

Data access and memory need a purpose, must be necessary, must be authorized and must have an expiry or review point.

8. ZERO-TRUST TOOLS

Every external tool, connector and agent capability is bounded by explicit scope, logging and containment. Prompt injection and untrusted content are treated as tainted inputs.

9. UNKNOWN IS A VALID ANSWER

Uncertainty is surfaced rather than silently converted into confidence.

10. FAILURE MUST TEACH THE SYSTEM

Failures, red-team results, rejected actions and corrections feed a failure atlas and evidence trail rather than disappearing from the record.