Prisme case study, part 2 of 3

Public spending

A section of Prisme that links each public scheme to what it costs, to the results observed among the people it reaches, and to the evaluations that measure its effect, with a level of evidence computed by a public rule.

Budget: €578.0bn spent by the State in 2025 on its general budget, across 33 missions and 135 programmes, each with its voted and executed credits and its performance indicators
Schemes: 29 public schemes, from the prime d’activité to the research tax credit, each with its cost, its reach and its evidence map
Evaluations: 106 studies cited, each with its method, its rank and whether its authors are independent
Documents: 206 documents kept with their fingerprint, so anyone can tell which version was read
Code: 10,700 lines of Python, 25 dbt models, 247 Python tests and 16 decision records, in a repository of its own

Counted on the live site and in the registry on 25 October 2026.

The question

What does a public scheme cost, and what does it achieve?

The answers exist, but they are scattered. The voted credits are in the finance law, the executed ones in the annual performance reports, the effect in academic papers and in reports by the Cour des comptes. They are rarely put side by side, and the strength of each piece of evidence is seldom said. A cost per beneficiary doesn’t measure efficiency, and the share of beneficiaries in work six months later doesn’t say what they would have done without the scheme.

The section started on 22 October as a project of its own, with its own name and site. The next day it became part of Prisme: two sites meant twice the legal pages, the data watch, the SEO and the branding, for one person.

The home of the public spending section: the headline “Le coût des dispositifs publics et les preuves de leur effet” and three tiles for the budget, the programmes and the schemes — For readers, it’s Prisme: same header, same footer, same trust pages. The section adds its own sub-menu.

One question per scheme

Each page asks one question in its title: what did it cost, and what do we know about its effect? Its first paragraph answers with the figures that are known, and says which version of the scheme each one is about.

The Garantie jeunes page: its question, a summary in which every figure is a link, and an identity card giving its status, funders and number of evaluations — Every figure in blue leads to its source. A scheme that no longer exists keeps its page, with its status.

The seven links of the Garantie jeunes, from means committed to attributable impacts, with the value of each link — Seven links, from the money committed to the impact that can be credited to the scheme. Two of them aren’t documented, and the page says so.

The state budget page: €578.0bn spent in 2025 against €582.4bn voted, and a bar chart of spending by mission — The state budget, mission by mission. Executed comes first, and the gap with what was voted is shown without being judged.

Four stages, never added

An amount is planned in the finance bill, voted by Parliament, opened during the year or executed. Two stages never add up, any more than commitment and payment appropriations do.

Four kinds of information

A published fact, a calculation by Prisme with its formula, an estimate by a third party, or an interpretation, always attributed. Each value on a page carries its kind.

A result is not an effect

Only an evaluation that builds a credible comparison group measures an effect; a model can estimate one, under its assumptions. A performance indicator doesn’t measure an effect either, and hitting a target says nothing about efficiency.

Nothing ranked

No scheme gets a score or a rank: effects measured on different criteria and populations don’t compare. A cost per effect is never recomputed; the page quotes the one an evaluation publishes.

A level of evidence, computed

Two things are shown apart and never merged into a score: the direction of the effect, and how solid the evidence is. A rule in the code computes the level for each criterion, never for a scheme as a whole.

The evidence scale, version of 23 October 2026. Ranks A to D describe a study’s method.
Level	Label	Rule
0	No evaluation of the effect found	No evaluation measures this criterion.
1	Observed results, effect unknown	Only descriptive studies (rank D): beneficiaries followed over time, before and after with no comparison.
2	Hints	One study with a rough comparison (rank C), a simulation, or a strong study (rank A or B) that is neither independent nor published with its data and code.
3	Probable effect	At least one strong study (rank A or B) that is independent, or whose data and code are published.
4	Established effect	At least two valid strong studies from different institutions, agreeing on the direction, one of them on the current version of the scheme.

The evidence map of the Garantie jeunes: four criteria in rows, three studies in columns, each cell with the effect measured, and the level of evidence in the last column — The evidence map. An empty cell means the study doesn’t measure that criterion.

What lowers a level

Strong studies that contradict each other on the direction of the effect.
A study of another version of the scheme, another population or another period.
An effect measured over less than 24 months when the official goal is long term.
An imprecise estimate, whose interval is compatible with no effect at all.

Two levels when the studies are about a past version

Schemes change their amounts and their public. A solid study of a past version doesn’t say what the current one achieves, so the criterion drops a notch, and the page shows beside it the level for the versions the studies cover. The rule also settles eight points the method left open; they stay proposals until researchers have reviewed them.

How a value gets into a page

Nothing is written at build time. A page assembles values read in documents, published data and declared calculations. Budget amounts and performance indicators come from published datasets, and their totals per programme are checked against a second one.

ArchiveThe cited document is downloaded and kept with its fingerprint. It is quoted, never republished, and treated as untrusted data: nothing in a PDF is ever run.
Read twiceEach value is read twice, independently, from the same questionnaire. Neither reading sees the other, and each works from its own text extraction of the document.
AnchorThe quoted sentence has to be on the cited page of the document, and the number in the sentence. The check is automatic.
CompareWhen the two readings differ, the value isn’t published until a person has decided, document in hand. The page says a decision is pending, so a missing value doesn’t pass for one that doesn’t exist.
VerifyThen a person checks each value at the cited page. That hasn’t been done yet, and the section’s method page says so, once.

The values and sources table of a scheme page: each value with its identifier, its kind, its document and page, the exact quote and the state of its two readings — Every value with its document, its page and the exact sentence. The whole registry downloads as CSV and JSON, with its schema.

One site, two repositories

For readers, it’s one site. Behind it, two builds that can’t break each other.

The section keeps its own repository, warehouse, evidence registry, CI and deployment. It builds its whole site under /depenses-publiques/, and Prisme’s Vercel configuration rewrites that path to it. The header, mega menu included, and the footer are the ones Prisme publishes at each build: the section fetches them at the start of its own, falls back on a versioned copy if it has to and says so, and a copy older than 30 days stops a strict build. Once built, every page is checked: links, canonical address, weight, Prisme’s header and footer, and the error report address of that page.

What’s left

The chain is in place. The human part comes next.

Settle every gap between the two readings, check each value at its page, have the texts of each page reread, and submit the evidence rule to researchers. Each page will also go to the administration and the operator concerned, for a factual check, with no right of veto; their answers will be summed up on the page.

Visit the section