The registry

The registry is the shared map of where to look. Each analysis ends. The registry stays, and gets better with each one.

Many analyses, one registry

Every analysis reads from the same registry and writes back what it learned. Nothing is copied. A fact learned about a source in one analysis is there for the next one the same day.

What it holds

• Source. One document: where it came from, its text, and a few lines on what it covers.
• Publisher. Who wrote it, who funds them, and whose side they are on. Every source has one.
• Who analyses are about. Ministries, parties, unions, officials: a name and the other names it goes by. Two analyses of one country share them. Anything true or false about them is a claim.
• Source type. A kind of document that comes out on a schedule, at an address we can predict: an IMF country report, an audit office's annual report, a census. It carries notes on using it well, such as "tables are scanned images before 2019" or "runs 14 months behind the fiscal year".

A source type works in any country, which is what lets the registry carry over from one analysis to the next. Most sources have no source type. News stories, think-tank papers and blogs are found by searching, so there is no route to record. For those, the publisher is what matters.

The registry says what a document is. It never records what the document says. Only an analysis reads a document for claims, and a claim belongs to the analysis that made it.

The analyst can promote a well-grounded claim, such as a budget figure, to the registry. Other analyses can then cite it instead of finding it again. A judgement is never promoted, because it is one author's reading. A claim from an interview is not offered for promotion unless the analyst picks it, and should be promoted only if the person it came from agrees. When a promoted claim is replaced, the analyses citing it should be told; that is not built yet. We are trying this; if no claim gets promoted in the first analyses, claims stay with their analysis only.

How it grows

Published analyses come first. An agent reads their endnotes and records what each author cited: the publisher, the document, and a source type where there is one. This shows what experts actually reach for.

Discovery then builds the registry for a country before the first analysis. It works through a checklist of institutions most states have: finance ministry, central bank, statistics office, audit body, parliament, the courts, the main newspapers. For each it asks what it publishes, where, and whether it can be reached without a login. Other agents look for datasets and for services worth connecting.

Each analysis leaves notes as it works. This route returned nothing. This PDF was a scan. This think tank turned out to be funded by the ministry.

Curation reads the notes. An agent drafts the fixes, a person approves them, and the change carries their name.

What an analysis may write back

Source types are kept few on purpose. If every analysis could add one, "IMF country report", "IMF Article IV" and "IMF staff report" would become three entries, and none of them would carry over to the next country. So a proposed type is first compared with the ones that exist, and an agent merges the near-duplicates that get through. A person still approves each type before it counts.