Stage 0 Research Project · Open for Public Critique

Proposed guardrails for AI in public decisions

Better public decisions, with AI under human authority.

Governments are starting to use AI to study complex public problems. NousPolis is building a public rulebook for using AI to research questions and stress-test ideas—while keeping human decision-makers in charge.

  • Citizens would raise community problems and share local evidence.
  • AI systems and human reviewers would gather sources, compare options, and challenge each other’s blind spots.
  • Elected officials and public institutions would make the final call and remain accountable to the public.

AI can research, compare, forecast, and challenge. It cannot authorize public policy.

Where we are today: NousPolis is an early-stage research and software prototype (Stage N0), not a government service. It is not used to make public decisions and has no public-policy authority.

Illustrative example

See the difference on a public question

Should our city allow taller apartment buildings near train stations?

Imagine residents debating rent, displacement, traffic, and green space. Here is the difference the proposed approach is meant to make.

One AI answer

A single polished answer can sound convincing while smoothing over disagreement, leaving out local concerns, or making it hard to check which sources mattered.

The NousPolis approach (proposed)

Residents could submit concerns and local evidence. Multiple AI systems and human reviewers would check sources, highlight gaps and uncertainty, and show competing options side by side. Elected city leaders—not AI—would decide whether to change the rules.

Illustration only: this is not a live policy case or a currently deployed service.

The hidden influence before a vote

Why keeping a human in the loop isn’t enough.

A human taking the final vote cannot protect democracy if an AI system—or the organization operating it—quietly chose the options and filtered the evidence first. Three questions matter:

Who gets the final vote?

Only public bodies with lawful authority can approve, change, or reject a policy.

Who shapes the menu of options?

Who writes the question, chooses the evidence, summarizes disagreements, and decides when research is finished?

Who decides when extra guardrails apply?

Who decides that a proposal is high-impact enough for stronger checks, independent scrutiny, or public hearings?

Real human control must cover the agenda and the safety rules—not just the last signature.

Technical spec: the three kinds of power

Technical terms: substantive / decision power; procedural / process power; definitional / rule-setting power.

Proposed

Who would do what?

Research can be shared. Responsibility cannot.

NousPolis proposes distinct roles so that better analysis does not turn into permission for AI to govern.

Residents and affected communities

Would raise questions, offer neighborhood knowledge, dispute missing evidence, and inspect the reasons for a decision.

AI systems and human reviewers

Would find and compare sources, test rival explanations, show uncertainty, and challenge weak conclusions.

Elected officials and public institutions

Would decide within their lawful powers, protect rights, explain their reasons, and answer for outcomes.

Stage N0 · Research prototype

Where the project stands

Where we are today—and why the next gate stays locked.

We are testing whether rules for evidence, review, and human decisions can actually be enforced in software. Passing a few tests is not enough to claim institutional readiness.

Today · Stage N0

Research and working prototype

On 17 August 2026, the core project merged an experimental, open-source policy-process pilot (v0.1). It tests whether software can check that evidence is recorded, analyses are challenged, and human decisions are documented. It has no authority to make government decisions.

Next gate · Stage N1

Locked until all checks pass

Stage N1 would allow only a non-binding research laboratory, not live policy decisions. The published safety check passes 9 of 11 grouped requirements, but independently witnessed records and review by genuinely separate people are still missing. The gate stays closed.

Technical status and dated evidence

The latest inspected N1 gate is dated 23 August 2026: grouped requirements A–K scored 9 of 11 passing, with requirements I (external integrity witness) and J (principal-separated review) not satisfied; the overall result was FAIL. Separate normalization-readiness checks also remained blocked. The experimental NP-PDLS-001 pilot is a test of process rules, not proof of external independence, evidence truth, policy correctness, or lawful deployment. It remains an N0 test instrument.

Read the exact N1 requirements and evidence →

Stewardship today

Founder-led, founder-funded, and disclosed.

NousPolis is currently founded and solely funded by Deon Van der Westhuizen. Founder control is a bootstrap condition, not the intended institutional end state.

Read the founder and stewardship disclosure →

Why this matters

AI can analyze thousands of pages. Who decides what it reads?

AI can help public agencies compare evidence and policies at a scale no single team could manage. But whoever chooses the questions, sources, risk labels, and stopping rules can quietly shape the outcome. NousPolis is designing rules to make that influence visible and challengeable.

Help test the idea

Bring a question, challenge an assumption, or inspect the rules.

Bring a local policy questionResidents, civic groups, and local leaders: suggest a real issue we could use as a non-binding research example.Suggest a public question →
Contribute researchShare evidence, comparable institutions, or policy-process research.Discuss research →