Skip to content
All writing

A constitution you can run: Building governable agents for real work

Funding matching shows why an agent needs explicit rules, evidence and a way to handle missing information. Model behavior is only part of that design.

Essay · AI · Governance · February 5, 2026 · 8 min read

Engraving: a craftsman winding a pinned music-box cylinder.

When an AI recommendation influences a budget, an application or an operational decision, someone will eventually need to explain it. Which information did the system use? Which rules applied? What did it do when the evidence was incomplete?

A fluent explanation can make a result easy to read. It does not, by itself, establish that the result follows from the evidence. I want the system to retain enough of its work that another person can check it and correct it later.

That is what I mean by a constitution you can run: explicit constraints, evidence requirements and controls that affect what the software does. The useful question is how those controls behave when the input is incomplete or two parts of the system disagree.

When an answer starts to influence a decision

A draft that someone has time and expertise to review allows room for mistakes. The consequences change when an output leads someone to commit money, choose a supplier, file an application or claim compliance. Even a summary can cross that boundary if a reader treats it as an authoritative account.

At that point, I would expect a system to help answer four questions:

  • Why was this option recommended?
  • Which sources support the material claims?
  • Which rule or missing fact could change the outcome?
  • What would we need to reproduce the check later?

Those questions also help establish who remains responsible. A recommendation can be useful while still requiring a decision from a person with the relevant authority.

Funding matching exposes the difficult parts

Funding matching can look like a search problem: compare a project with a program and estimate the fit. Thematic similarity is useful, but it leaves out much of the decision.

A project can fit a program’s purpose and still fail its eligibility requirements. Region, company size, cooperation partners, technology readiness level, application dates, funding caps and the treatment of particular costs can all matter. Combining programs may introduce further limits.

There are also cases where the answer is not yet knowable. An applicant may not have supplied the relevant budget detail, or a program’s wording may need interpretation. I want the system to distinguish an ineligible project from one that cannot yet be assessed. Treating both as a low score would lose information the adviser needs.

My preferred approach is to search broadly enough to find plausible candidates, then narrow them through explicit checks. A recommendation should include the evidence and any unresolved questions. When a necessary input is missing, the workflow should ask for it.

Separate thematic fit from eligibility

Semantic retrieval and ranking help find programs whose language differs from the project description. They belong in discovery, where missing a plausible option is costly.

For explicit eligibility rules, I would use separately testable logic. It can check thresholds, exclusions, caps and combinations once the relevant facts and rules have been formalized. The formalization itself needs review: a deterministic engine will consistently apply a wrongly encoded rule just as readily as a correct one.

A merge step can bring the signals together, but it should preserve disagreements. High thematic fit with failed eligibility is a useful result, provided the failed condition is visible. Passing eligibility with weak thematic fit may call for more project context. A single weighted score should not conceal either situation.

Some conditions resist a clean formal rule. Those need a documented interpretation or a review route. The purpose of separating the checks is to make their limits visible and their behavior testable.

What model constitutions contribute

Anthropic’s January 2026 constitution for Claude describes the behavior it wants from the model and explains the reasons behind its priorities. Anthropic uses the document in training, including to help the model handle unfamiliar situations and conflicting objectives.

I find that useful because it makes the provider’s intentions available for inspection. Anthropic also states that outputs may fail to meet those intentions. A product team still needs to decide which actions the model may take and how to enforce the limits of its particular workflow.

For example, a model’s instruction to avoid fabrication does not replace a check that a recommendation cites an available source. A general preference for helpfulness does not establish permission to submit an application or change a record.

A paper by David Peer and Sebastian Stabinger of Otera describes Autonomous Trustworthy Agents, or ATA. Published in October 2025, it separates the formalization of a knowledge base from the processing of individual inputs.

In their insurance-claims example, an LLM translates policy terms into a formal representation that experts can inspect and correct. Incoming claims are encoded separately, and a symbolic engine evaluates them against that knowledge base. The authors report competitive results on their evaluation task, with improvements after human correction of the knowledge base.

The separation is useful to me because it creates a place to inspect rules before relying on them. It also narrows the part of execution that can be deterministic. The language-model steps and the correctness of the formalized facts still need evaluation. Results on one claims-processing task do not establish that the architecture will make every decision workflow reliable.

For funding matching, I would apply the same separation between interpreting program documents, approving structured rules and assessing a particular project. That is an architectural choice to test against the domain’s requirements.

The controls I would build around the workflow

A versioned catalog provides a reference for the programs and source documents used in a run. When a program changes, the system needs to distinguish a new assessment from the reasoning behind an older one.

Structured intake captures the facts that eligibility depends on. A free-text description can help with discovery, but it should not be the only record of company size, location, partners or the proposed budget. Missing fields need an explicit state.

The rule engine should have its own tests and version history. The merge logic needs documented weights and a way to expose conflicts. The explanation step should work from those recorded results and sources, so that it can describe the assessment without inventing a rationale afterward.

I also want a record of the inputs, retrieved material, rule results, model and prompt versions, and final recommendation. That gives the team a basis for investigating a decision. Re-running an external model does not guarantee identical output; where exact reproduction matters, retaining the original intermediate results is necessary as well.

The interface can support this by showing how a candidate moved through search and checks, which requirements remain unresolved, and where a person needs to decide. Users should be able to inspect the important conditions without reading an internal execution log.

Evaluate retrieval, decisions and explanations separately

I use three questions as a checklist when reviewing these systems.

Did we retrieve the right material?

Check corpus coverage, source authority, freshness and access permissions. If the program document is missing or out of date, later stages may produce a well-formed assessment of the wrong information. Retrieval tests should include known relevant sources and examples the system is likely to overlook.

Did we apply the rules correctly?

Test explicit constraints independently of the language model. Include boundary values, exclusions and combinations of rules. Record which version applied to a decision so that a change can be reviewed and its effect understood.

Does the explanation match the evidence?

Check that citations support the claims made and that the text distinguishes confirmed facts, interpretation and missing information. A useful explanation should also tell the reader what could change the assessment.

These checks can fail independently. Keeping the results separate makes it easier to find and repair the problem. Combining everything in one model call may be convenient for a prototype, but it makes this kind of diagnosis harder.

Match the controls to the consequences

Maintaining a formal rule engine, a versioned catalog and detailed execution records takes work. I would not apply that whole design to every drafting or exploratory tool.

The level of control should reflect the data involved, the consequences of an error and how reliably a person can review the output. A human approval step helps only if the reviewer has the information, expertise and time to make a meaningful judgment.

It is also worth revisiting those assumptions when the product’s use changes. A tool introduced for informal exploration can become part of a consequential workflow without anyone explicitly deciding that its controls are sufficient for the new role.

There are legal requirements to assess as well. The EU AI Act includes obligations for defined categories of systems and operators. Its Article 99 distinguishes penalty categories: the €35 million or 7% ceiling concerns prohibited practices, while specified other obligations fall under a €15 million or 3% ceiling. The applicable obligations, dates and penalty rules depend on the system and operator. A generic claim that an agent is “governed” does not establish compliance.

Let evaluation change what happens next

Evaluation is most useful when a result has a consequence. Missing evidence can prevent a recommendation from being presented as settled. A conflict between thematic fit and eligibility can trigger a review. A required fact that is absent can lead to a request for information.

Those responses need to be designed, tested and assigned to someone. Otherwise the team may have a dashboard showing recurring failures without a practical way to address them.

I would increase an agent’s freedom as the evidence supports it. Reversible, bounded work can proceed with lighter review. More consequential actions need stronger evidence and clearer authority. The record should make it possible to understand why the system was allowed to act in a particular case.

For funding matching, that means being able to show an adviser the candidates found, the conditions checked and the questions left open. It gives them something they can examine and use in a decision. That is the standard I would build toward.