Our method, explained.

Hard rule. A language model is never used to produce an amount. Its role is limited to text: categorising, describing, finding an address, writing a title. All of the site's financial aggregates come from reproducible calculations, run directly on raw data. The texts it writes are: the assistant's answers; the plain-language descriptions of Paris contracts, projects and grants (the “In practice” paragraphs); some of the readable titles of towns' public contracts (the others are rewritten by rules, without a model); the “What it does” sentences of associations, drawn from the purpose they declared; the titles of the Council of Paris moments on place pages; the plain-language descriptions of San Francisco and Recife contracts, and the titles of Brazilian municípios' contracts.

Public sources, credited: every figure comes from a file or a document published by an administration, and the page that shows it says which. For the few Paris debt ratios the City does not publish as open data (weighted average rate, maturity for example), values are marked “indicative” on the relevant panel.

Individuals: aggregated, never by name. The administrative-account annex also lists aid paid to individuals — a few hundred euros each, labelled “Personnes physiques” by the City itself. The source dataset publishes them by name; this site makes a different choice: amounts remain visible as a per-year aggregate, but no individual's name is indexed or searchable. A name attached to an aid payment reveals a person's social situation, with no public-scrutiny value at the individual level — these rows carry neither purpose nor department. Some years (2020-2021) already arrive aggregated as a single line in the source. The name-level detail remains available in the City's dataset. The grant lists of the other towns do not show any private individual either.

Tracing a figure, tool by tool

Click a tool to see where its figures come from: each step of processing, all the way back to the public dataset they are drawn from.

Freshness by source

Coverage by tool

Rule: the current year is always provisional until the final administrative account is published (~June N+1). Figures for an unclosed year are based on the voted budget.

Automated checks

For Paris, on every update, automated checks replay the calculations and match the displayed totals against the raw sources, to the cent. Model-enriched information (categories, locations) must stay above documented reliability thresholds. And any gaps in the public sources are flagged as such on the relevant page. No figure is typed by hand.

Loading audit…

Known limitations

The model rates itself. When the language model proposes a category for a grant or a location for a project, it attaches a confidence score to its answer. That score is self-declared. To make it reliable beyond this self-evaluation, it will need to be compared against a hand-labelled sample. That work is underway.

When our figures differ from the City's. On a given dataset, our totals are strictly identical to the City's: we never alter an amount. When a gap shows up, it usually comes from one of three reasons. Either the scope differs: we publish the main budget alone, while the City may communicate a "City group" that includes its satellite entities. Or our figure comes from the latest published dataset, itself sometimes lagging an official communication. Or we have translated a label without the underlying aggregate changing. If a gap persists without an obvious explanation, it's probably a bug. Report it.

The code of the pipeline that computes these figures is published under the MIT licence: you can read in it how each figure is obtained from the public sources. The code of the website and of the enrichment steps is not published. When an error is reported and confirmed, it is fixed in the calculation. Significant corrections are published on the Corrections page, with their date and where the report came from.