Method

Status: working version, October 2026. Licence: CC-BY 4.0.

The catalogue documents AI and other automated decision systems and the people they reached. It holds two kinds of record: in-depth case-files and shorter registry records.

Inclusion criteria

A system is included when all of the following hold:

1. Deployed. It was used in practice, not only planned, piloted on paper or proposed.
2. Automated output about people. It produced a score, flag, match, ranking or decision about specific people or a defined group — whether by machine learning, statistical models or fixed rules.
3. Documented consequences. That output led to documented consequences for those people: a decision, a stop, a removal, a denial, a report to authorities, a loss.
4. Public record. The above is documented in public sources that meet the sourcing policy.

Not included: surveillance capability with no documented decision about people; spyware and hacking; human misuse of databases; systems that were never deployed; policy arguments without a documented case. Excluded records may be reconsidered if new evidence appears.

1. Case-files (S-NNN)

A case-file is a full forensic record of one system, built by hand from public sources to a fixed evidence standard:

• at least three primary sources (court rulings, regulators' decisions, parliamentary or official inquiries, statutes, peer-reviewed audits) and at least three secondary sources;
• a claims matrix in which every critical claim is supported by at least two independent sources, or is marked as single-sourced or contested;
• archived copies of sources where retrieval allows.

Each case-file codes harm on seven dimensions — liberty (LIB), dignity (DIG), employment (EMP), family (FAM), housing (HOU), health (HEA) and reputation (REP) — each from 0 to 10, using the operational rules summarised in Coding rules. Each score carries separate marks for evidence confidence and causal certainty.

• The dimensions are never added together. There is no total score and no ranking of cases.
• Where evidence is too thin to score, the cell says so: INS (insufficient) or SUSP (suspected, documented at population level but not measurable). These are not zeros.
• Each system is classed as A (machine learning / AI), B (statistical model, or hybrid scoring composed with human decisions) or C (fixed rules). Class C is not a lesser category: rule-based systems have caused some of the most severe documented harms.

Case-files are re-coded when the facts change. Each record shows its version and coding date.

2. Registry records (R-NNNN)

Registry records are shorter, structured entries found by an automated collection pipeline and then checked by a person. The pipeline is described in Collection methodology and its prompts in Extraction prompts. Only records a person has verified are published. Records still in review are reported as a count only.

A registry record is not a case-file. It is a documented lead with sources; some registry records may later become case-files.

Ethics

• Only public sources. No leaked or breached data, no content from closed platforms.
• People harmed by a system are named only if they chose to make their own case public. Otherwise they are described, or identified by initials.
• Children are never identified. Health harm in cases involving deaths is recorded at aggregate level only.
• Living people connected to a case can ask for a correction or for their information to be removed: see Corrections in the colophon.

Limitations

• Coding is done by a single coder; inter-coder reliability has not yet been tested.
• The seven dimensions do not capture every kind of harm.
• Well-documented cases are over-represented: harm that leaves no paper trail is harder to record. The catalogue marks how well documented each system is rather than treating documentation as a neutral fact.
• Scores are structured judgements, not measurements. They cannot be used to infer statistics about populations.

Documents

• Coding rules — summary of the Severity Index Codebook v0.5 (working version)
• Full Methodology and Codebook — forthcoming with v1.0