Styxis
Company
Knowledge

Count authorities and departments separately in a local-government AI pilot

A new-use experiment needs a defined experimental unit before its application count can be interpreted.

Choose the experimental unit before counting

Growing applications can look encouraging when a new AI-service pilot is being prepared. Mixing authority and department counts nevertheless obscures where usefulness will be tested. Begin with a small counting exercise on the application register. Its success criterion is an explanation of why totals differ under different units, rather than the largest possible number of participants or an apparently impressive scale of adoption.

Recruitment is not a pilot result

Japan’s Digital Agency announced recruitment for a national-local government AI-service pilot on 20 October 2026. It seeks roughly 50–100 authorities and allows several departments from one authority to participate. The trial begins later. A larger department application count should not be rewritten as a larger number of authorities that have already completed a successful experiment.

Build two views from one register

A proposed working register retains authority identity, department identity, receipt time, eligibility state and pilot stage. Produce an authority view and a department view separately. Two departments applying from one authority should remain one authority in the first view. Recruitment staff must confirm what institutional boundaries and department changes mean; software cannot infer those organisational identities from names alone.

Select a narrow Truthound check

Truthound Core Python SDK 3.3.4 (open-source SDK, Polars backend) documents MultiColumnUniqueValidator for caller-defined composite keys, such as an approved authority-department key. DistinctCountValidator can check an externally defined cardinality range. The owner specifies the keys and expected bounds; the recruitment target should not automatically become a pass threshold. Neither check determines eligibility or establishes that a pilot produced useful public-service outcomes.

Test amendment against separate participation

An amended application from the same department may not be a new participating department. Another department in the same authority may genuinely run a separate experiment. Do not finalize selection counts while the register cannot distinguish those cases. Explain amendments and counting units first. Splitting an application into several supposed outcomes would make the growth look stronger while reducing the usefulness of the pilot record.

Keep an existing correct key check

pandas 3.1.1 (general-purpose DataFrame library with project-owned checks) DataFrame.duplicated can flag records using an explicit subset and keep policy. Retain a recruitment workflow that correctly handles its mask and amended applications. Consider Truthound when named key and expected-range checks would help the handover criteria. The interface choice does not decide the population’s meaning, and capability documents do not show that either product improves recruitment quality or trial effectiveness.

Define the small experiment’s failure condition

First examine whether the register distinguishes several departments within one authority from an amendment by one department. This article reports no execution on real applications. Hold the aggregation if identity remains unclear or application and selection stages are mixed. Explaining the resulting difference is a prerequisite for extending the same approach to regional comparisons without changing the unit halfway through the analysis.

Ask what comes after the count

Once units are clear, define which workflow will be tested and over what period. Counts of authorities and departments are inputs to that design, not evidence of usefulness. A well-managed application register should make the next small experiment’s observations and stopping criteria explicit. Its purpose is to support a credible test, rather than make the adoption programme appear larger before results exist.

Conditions

Separate authorities from departments and define application, selection and pilot stages.

Alternatives

pandas 3.1.1 duplicated can retain checks on an approved authority-department key.

Limitations

  • Several departments are not several authorities; application is not participation or demonstrated benefit.

Evidence

Event checked
Subject checked
Modified