A condition of the ambition
AI governance.
Evidence.
Human authority.
Institutional and sovereign-scale reasoning needs more than persuasive answers. It needs explicit authority, inspectable evidence, meaningful uncertainty, and control that remains with people.
THE NORTHSTAR SCIENTIFIC CONTRACT
The answer is not enough.
Show the work.
NORTHSTAR is being designed to return an inspectable evidence path. The question, the method, the calculation and the limits belong with the answer.
Problem. What would count as an answer? Keep the requested result, units and acceptance conditions attached to the inquiry.
Assumptions. Expose what the method assumes. A change in the initial conditions or the scope should be visible in the resulting evidence.
Derivation. Preserve the formulation and the steps that connect it to the calculation. Models may propose a method; the mathematics remains inspectable.
Execution. Record what ran, with which inputs and configuration. An execution receipt identifies an attempt; a successful process alone does not establish a correct result.
Verification. Compare the result with declared tolerances and independent approaches. Preserve failed checks and disagreement rather than smoothing them away.
Uncertainty. Publish the supported conclusion with its uncertainty. Inconclusive, contradicted and insufficient-precision outcomes remain meaningful outcomes.
No material claim without a traceable derivation or compute receipt.
Illustrative evidence contract · select a stage to inspect it. No model or worker is contacted.
03 / RETAINED AUTHORITY
Connected intelligence. Deliberate boundaries.
Identity, data, policy, and evidence stay inside explicit boundaries. Human purpose remains at the center.
Human purpose, not autonomous mandate
People and institutions define the mission, its limits, and the decisions it may inform. NORTHSTAR’s executive is intended to coordinate work inside that authority—not invent new authority or substitute for democratic, scientific, professional, or institutional judgment.
Cross-evaluation, not majority voting
Contributors should examine sources, challenge assumptions, expose contradictions, and test claims. Agreement among correlated models is not independent confirmation. The system must be able to say that evidence is missing or a disagreement remains unresolved.
Evidence that can be examined
Results should bind to their mission, model or tool identity, inputs, execution attempt, and verification status. The goal is traceable findings, citations, allowed explanations, and uncertainty—not collecting or exposing private hidden model reasoning.
Sovereignty through explicit boundaries
Data residency, tenant isolation, identity, credential ownership, infrastructure access, and external disclosure all need explicit policy. “Sovereign” is not a claim of automatic compliance or a label that removes the need for independent assessment.
Less authority than capability
A system may be technically capable of reaching a provider or changing a network without having permission to do so. Registration is not admission. A model’s output is not approval. Every consequential effect needs its own scoped execution boundary.
Honest recovery
A timeout is not proof that an external action failed. Cancellation is not proof that it stopped. Ambiguous outcomes should remain explicit until reconciled; they must not silently become permission to repeat the action.
Independent network evidence
Network systems remain external arms with explicit, versioned boundaries. NORTHSTAR never mutates network nodes, signs network commands, determines effective network state, alters evidence, or promotes Network Fabric claims. The optional Network Fabric lane is disabled by default.
Research claims must be earned
The vision is to advance synthetic reasoning for the benefit of humankind. Demonstrating that advance requires meaningful baselines, reproducible evaluation, adversarial testing, and evidence tied to the exact configuration under test.
This page is not a security certification, an assurance of production suitability, or evidence that every intended behavior is implemented.
Report a concern
Contact hello@mainelycode.com with a sanitized description and the relevant release or page. Do not include credentials, confidential research, or private customer data in an initial message.
Clear answers
Questions about the architecture
Do specialist models vote on NORTHSTAR’s final answer?
No. Specialist regions advise. The executive core arbitrates the evidence and maintains the final publication boundary. Model agreement alone does not establish truth or authorize an action.
Does a private deployment automatically establish sovereignty?
No. Sovereignty requires explicit control of data movement, identity, infrastructure, policy, and operational ownership. The relevant deployment must be independently assessed; a location label is not certification.
Can NORTHSTAR control the Network Fabric?
The optional Network Fabric lane is disabled by default. NORTHSTAR never mutates network nodes, signs network commands, determines effective network state, alters evidence, or promotes Network Fabric claims.