The protocol
PuroChile is not a way of finding figures. It is a reconciliation protocol: the rules under which AI agents bring any number of sources —each in its own format, each counting a slightly different population— to one result, and do not publish it until it has been checked against what every source published. Chile is the first synthesis made under it.
15 official sources · 1,144 checks · 56,273 published figures
Running
last run · 2026-10-17- 01read15 sources · one reader per source
- 02decompose% → count · average → weighted total
- 03sumcommune → province → region → country · rate at the end
- 04reconcile1,144 checks against what each source published
- 2024 census · INECensus populationpublished 18,480,432 · rebuilt 18,480,432exact
- Casen 2024 · Ministry of Social Developmentsuma de regiones (expr)published 20,131,682 · rebuilt 20,131,682exact
- Casen 2024 · Ministry of Social DevelopmentIncome povertypublished 17.3 · rebuilt 17.29rounding
- Casen 2024 · Ministry of Social Developmentsuma de comunas (expc)published 20,131,682 · rebuilt 20,095,516 · Δ -36,166declared
- Tax register · SIIr01 empresaspublished 26,326 · rebuilt 26,326exact
- Ministry of Education · SIES2006 egresados jóvenes desde comunaspublished 205,792 · rebuilt 205,792exact
- Hospital discharges · DEISegresos r01published 28,847 · rebuilt 28,847exact
- Agricultural census 2021 · INElos 15 usos suman la superficie censada (Chile)published 45,742,565 · rebuilt 45,742,565exact
- 2025 elections · TRICEL and SERVELpresidencial, primera vuelta: mesas cuyo total no cierrapublished 0 · rebuilt 0exact
- Foreign trade · Aduanasexportaciones enero de 2025: total del mes (US$)published 9,630,635,004 · rebuilt 9,630,635,004exact
- 05publish975 exact · 49 within rounding · 120 declared → site · MCP
The problem
Finding the data is the easy part. The hard part is putting it together: an agent that synthesises many sources does not fail loudly, it hands back something plausible. Four ways a synthesis of Chile comes out flawless and wrong:
A commune of 2,000 weighs as much as one of 600,000, and the result has exactly the shape of a national percentage.
Casen expands to about 20.1 million people and the 2024 census counted 18.48. A share from one times a count from the other gives a figure with all its decimals and no meaning.
Within a single tax-service release, one file writes 27.849.823 and another 133727798.94. With the wrong decimal mark, millions become decimals: six orders of magnitude, and the figure still looks reasonable.
Casen leaves the commune of origin blank for people who never moved. Comparing origin with destination left every one of them out: a flow diagram that balanced against itself and in which nobody had stayed. It happened here, and a check against the population caught it.
Six rules no step skips
The agent does not write figures
An agent reads, decomposes, sums and compares; every cell comes from a file a Chilean public authority published. What the source does not publish, the synthesis gives as “no data”: nothing is interpolated or estimated to fill in a map. The two figures that do come from a model —the projections of the tax register and AI adoption by sector— are labelled where they appear and feed no indicator.
Sum first, divide last
Averaging percentages across communes of different sizes is the mistake an agent makes most naturally. So every source is decomposed before it meets another: a percentage goes back to the count behind it and an average to a weighted total. Only those are summed, and the rate is computed at the end.
Every result is reconciled with what was published
The national total is not copied: it is rebuilt from the 346 communes and compared with the published one. All 1,144 comparisons are published, including those that do not come out even.
A difference is declared, not corrected
Where the reconstruction disagrees with the source, PuroChile shows both figures and explains why. No step may adjust a number to make it fit.
Two sources are not crossed silently
Casen expands to 2017-census projections and reaches a population about 9% larger than the 2024 census counted. Its shares are never multiplied by census counts, and the few figures that divide one source by another take both from the same year.
A small cell names nobody
Where a fine figure could point at a handful of people —the programmes a small commune's school leavers go on to, a cancer site with five cases— it is grouped or withheld, and the page says so rather than leaving a gap.
How rates are computed
This is the error the protocol exists to make impossible: the step that adds accepts counts and weighted totals and nothing else, so a percentage cannot reach it. The percentage comes back at the very end, as a ratio between two measures that were summed first.
Source by source
Each authority publishes in its own format and each format has its own way of going wrong. Before any agent puts two sources together, each one is brought to the same shape — measures that can be added — and this is what that takes for each.
2024 census · INE
The census layers and the age-by-sex table are read. Each layer's identity is verified on every read, and where INE's service offers an aggregation, its own is used.
142 checks →Casen 2024 · Ministry of Social Development
The survey is expanded with the factor that belongs to each level — commune, province or region — and never with another's. A cell with too few respondents is not published.
40 checks →Tax register · SII
Each file's encoding and decimal mark are detected from its contents, and a commune name that does not match the census stops the read instead of being dropped.
106 checks →Ministry of Education · SIES
Two school registers are joined on the masked RUN Mineduc publishes in both, and each leaver's origin is the commune where they went to secondary school.
420 checks →Hospital discharges · DEIS
Each commune is standardised by its own age and sex before it is compared, and the year of a region whose register stopped reporting is withheld.
54 checks →Agricultural census 2021 · INE
Hectares are summed and holdings never are: a holding with land in two communes is counted in both, and is not split.
24 checks →2025 elections · TRICEL and SERVEL
The result the Tribunal qualified is summed, polling table by polling table, and not election night's count, which is read only to measure how far the qualification moved it. Turnout divides by the electoral roll and never by the census, and votes cast abroad are assigned to no commune.
89 checks →Foreign trade · Aduanas
The export and import declarations Aduanas publishes every month are summed, filtered by the operation types Aduanas itself names, and checked line by line against its tables by country and tariff code. Countries, customs offices and regions are matched by code, never by name.
87 checks →Census blocks · INE
Blocks and rural entities are aggregated into census districts, the finest level PuroChile draws. From each one the straight-line distance to the nearest health facility and school is measured: a floor, never the travel distance.
3 checks →Official cartography · INE
Boundaries are projected and simplified once, and oceanic communes are left out of each map's frame by a rule, not by a list.
Where to check it
1,144 checksEvery one of those checks is published with the figure the authority gave, the figure PuroChile rebuilt from the territories underneath, and the difference between them — including the ones that did not come out even.