State OS

State OS is a research project on decision provenance: keeping, for each decision a government takes, a checkable record of what was decided, on what evidence, against which alternatives, and with what expectations. It is designed to run in shadow mode, which means beside a real government, using only public information, with no governmental authority and no role in any decision.

"State OS" is a project name. It is not a claim that a state can or should be run by software.

The problem

Governments publish a great deal: decisions, statistics, legislation, parliamentary records, economic projections. In the Netherlands, for example, the Council of Ministers publishes a list of its decisions with a brief explanation of each [1].

What is rarely kept is the link between those pieces for one decision. A reader who comes back two years later usually cannot recover:

• the evidence as it stood on the day, because statistics are revised;
• which alternatives were weighed, and under which values each would have been chosen;
• what the decision was expected to achieve, in a form that can be scored;
• whether the analysis was written by people, by AI models, or by both;
• how the outcome compares with the expectation.

The fourth point is becoming more important. In 2023 the Associated Press reported that a Brazilian city council had passed an ordinance drafted with ChatGPT without being told [2].

Without a record fixed before the outcome, any claim that "we would have done better" can be made in hindsight and cannot be checked.

Main components

Decision Record. The basic unit. One append-only, time-stamped record per official decision. It contains the official text, a frozen evidence snapshot, constitutional flags, alternatives, dissent, forecasts, the human editor's review, and later the scored results. Every statement carries a label that says what kind of claim it is and whether it can ever be scored: fact, forecast, simulation, counterfactual, normative judgment, legal flag, or provenance record.

State Git. The provenance layer: a signed, mirrored, externally time-stamped repository that holds the official sources and every record. It is a ledger of what was recorded and when. It is explicitly non-authoritative; a merge means "recorded", never "adopted". Its promise is that rewrites can be detected by outsiders. It does not promise that they are impossible.

State Data Platform and scenario models. The intake for official data, with source, version, retrieval time and hash for every item. The project deliberately avoids the term "digital twin", because no validated socio-economic model of a whole state exists.

Shadow Analysis Cabinet. A set of analytical AI agents: six organised by policy area and five by function (coordination, forecasting, dissent, constitutional checking, audit). They are a way of dividing analytical work. They are not ministers and represent no one.

Authored Value Frameworks. Alternatives to the enacted decision are written under four declared frameworks: social-democratic, liberal, conservative and technocratic. Each is a versioned document with named authors and outside reviewers. They are authored position models. They are not the views of any party or group of citizens.

Constitutional Clerk. A narrow rule-checker over the official constitutional text. It can produce only three outputs: no flag on the encoded rules, a flag with a citation, or "human review required". It never says "constitutional" or "unconstitutional".

Forecast ledger. Probability forecasts about observable events on the path the government actually took, registered before the outcome and later scored beside naive and official baselines.

How they connect

Official sources are mirrored and hashed. A snapshot of the evidence is frozen and time-stamped. Every agent and every framework works on that same snapshot. Their outputs are assembled into a machine draft. A named human editor accepts, modifies or rejects each section and gives reasons. The approved record is signed, time-stamped and mirrored. When the forecast dates arrive, a separate scorer appends the results. Nothing is edited in place: corrections are new entries.

Use scenarios

These are intended uses. None is in operation.

• A journalist or researcher checks, a year on, what a decision was expected to achieve and whether the forecasts held, with every forecaster scored on the same questions.
• An auditor or oversight body verifies that a record has not been altered since it was made, and sees which evidence versions and which models were used.
• A reader comparing options sees the enacted decision first, then alternatives of equal prominence, each tied to a declared framework, with the strongest objection to every option including the enacted one.

Agent authority and the human role

The agents have no authority. Their outputs are inputs to a human-edited record and are never decisions. The coordinating agent cannot silently edit what the others produce; the raw outputs are kept. Agents that read untrusted documents have no tools and no write access. The dissent agent and the Constitutional Clerk run on different model lineages from the agents that propose options.

Nothing is published without a named human editor. This follows the EU AI Act, which requires disclosure of AI-generated text published to inform the public on matters of public interest unless it has undergone human review and a person holds editorial responsibility [3]. The project's choice is to do both: disclose and name an editor.

The design does not assume that a human signature means human judgment. It measures it: every accept, modify or reject is logged with a reason, the machine draft and the published text are both hashed so the extent of editing is visible, harmless errors are seeded into drafts to see whether they are caught, and a sample goes to a second reviewer.

Control, audit and limits

• Fixed coverage rule. Which decisions get a record is decided by a published mechanical rule, with every skip logged, so the operator cannot choose only the convenient cases.
• Separate scorer. Forecasts are scored with open, deterministic code by a person who does not configure the forecasting.
• Forbidden outputs. No single recommended option. No "win rate against the government". No statement that an alternative "would have been better". The effect of a policy that was not enacted is never observed, so such claims cannot be checked.
• No personal data beyond the names of public officials acting in office, and no outputs about individuals.
• Not a voting aid or campaign tool. The EU AI Act treats systems intended to influence voting behaviour as high-risk [3].

The limits of these controls are stated in the project's own documents. With a small founding team, "two approvals" means two of the same people. The founders' control would be visible and forkable, but it would not be plural. Nothing technical prevents someone from citing a record to lend neutral-looking support to a decision already taken. And the alternatives, the most distinctive output, have no source of legitimacy beyond their declared authorship: they cannot be validated by outcomes, and nobody authorised them.

Stage of the project

On 2 October 2026 the prototype ran end to end on eight real decisions from Dutch Council of Ministers lists of late August and early September 2026. According to the run output, it produced eight draft Decision Records, and each passed the project's automated checks. These are draft outputs of that one run. The checks test that a record has the required structure and is internally consistent. They do not test whether its legal flags are correct, whether the analysis is any good, or whether such a record would work in practice. Specifically:

• no draft has been read by a human editor, and no editor holds the role;
• no Decision Record has been published, mirrored or time-stamped in any public log;
• all four value frameworks are unreviewed drafts with no named authors;
• every draft was flagged unstable, because repeated runs on the same inputs gave different text;
• no forecast has resolved and no scoring has been done.

Data and model providers in that run. The inputs were published documents: the decision texts from the Council's lists, the constitutional text from the official repository, and aggregate statistics from the national statistics service. No personal data were used beyond the names of ministers as they appear in those documents, and no synthetic decisions. The text was generated by hosted commercial models: a Yandex model for the portfolio analysis, the alternatives and the forecasts, and Google Gemini models on the free tier for dissent. They were used because they were available on the day. The specification names a different configuration, so the run says nothing about whether using several model families helps. Whether a record of Dutch government decisions should depend on a provider based in Russia, or on a free-tier service, is a governance question the project has not decided. The concept paper gives the detail in section 12.

The project has no partnership or affiliation with any government and no government or party funding. A first public version would be judged against seven objectives fixed in advance, and the project has committed to publishing a null result.

Conclusions and practical implications

• The useful unit of accountability is the single decision. Most of the material already exists; the binding does not.
• Evidence snapshots and forecasts are only worth having if they are fixed and witnessed before the outcome.
• An automated legal check should flag and cite. It should never rule.
• People are the binding constraint. The code was the quick part.

Limitations and open questions

This page condenses the concept paper. The three external facts it cites carry numbered sources, opened on the date given below. The text was drafted with AI assistance (Claude, a model made by Anthropic), and the author is responsible for what is published. No one independent of the project has reviewed it, and it is not peer reviewed. The prototype results are the project's own and unreplicated.

Open questions that only operation can answer: how much editor time a record needs; whether editors' judgment holds up over months; whether people who hold a tradition recognise the framework written for it; whether the Clerk ever falsely reassures; whether anyone outside actually verifies the log; and whether readers mistake the system for an authority.

The full argument, the architecture and the evaluation design are in the concept paper: State OS: Concept and Architecture.

Sources

All opened on 4 October 2026 and checked again on 5 October 2026.

1. Rijksoverheid. "Besluitenlijst ministerraad 17 april 2026." https://www.rijksoverheid.nl/documenten/2026/04/17/besluitenlijstministerraad17april2026
2. Associated Press (Jeantet and Savarese), via KSAT. "Brazilian city enacts an ordinance that was secretly written by ChatGPT." 30 November 2023. https://www.ksat.com/tech/2023/11/30/brazilian-city-enacts-an-ordinance-secretly-written-by-a-surprising-new-staffer-chatgpt/
3. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50(4) and Annex III point 8(b). https://publications.europa.eu/resource/celex/32024R1689