How the library works
Evidence, with its limits intact.
We document applied AI as research, not promotion. A recognizable publisher can help; it never substitutes for provenance, methods, access, and corroboration.
What qualifies as a case?
A concrete AI intervention, incident, organizational decision, or economic change with identifiable actors and real-world stakes. Cases may be positive, negative, mixed, or inconclusive. A product announcement without observed consequences is a lead, not a case.
Sources are evaluated, not whitelisted
We welcome accountable journalism, primary records, regulators, scholarship, audits, technical postmortems, and independent investigators. Every core source is assessed for identity, access, method, provenance, independence, corroboration, accountability, and integrity.
Independent work may be the strongest available source. Its access conditions still matter: a group can be editorially independent while relying on evidence supplied by the organization under study. We disclose both facts.
Evidence grades
| Grade | Interpretation |
|---|---|
| A | Triangulated primary and independent evidence, with methods or data sufficient to assess major claims. |
| B | Good evidence and meaningful corroboration, with incomplete access, limited replication, or material unanswered questions. |
| C | Useful but constrained evidence that is largely self-reported, indirect, or only partly corroborated. |
| D | Lead-level material such as rumor, promotional claims, or unresolved provenance. Not publishable. |
Claim labels
A grade describes a whole case; labels describe single claims. Every material outcome, scale, cost, causal, harm, or forecast claim carries one, together with the evidence behind it and what would change it. A case can be well supported overall and still rest on one claim nobody has verified.
| Label | Meaning |
|---|---|
| Verified | Directly supported by strong evidence and corroborated where reasonable. |
| Supported | Credible evidence supports the claim, but proof, precision, or causality is incomplete. |
| Attributed | A named participant made the claim; the library has not independently verified it. |
| Disputed | Credible sources materially disagree. |
| Inference | The case author derived the conclusion from cited facts. |
| Unknown | Available evidence does not resolve the question. |
Which half of the error rate is missing?
Cases in this library from healthcare, manufacturing, and retail independently turn on the same omission: an organization knowing or publishing one side of its error rate and not the other. A vendor reports how many cases its model catches; nobody reports how often its alerts are wrong. That recurrence is a finding about the field, not a coincidence of what we chose to write about.
So every case involving a system that flags, scores, or classifies asks the same two questions. Which half is missing—and who bears the cost of the error that went uncounted? A false negative is usually absorbed by the organization. A false positive lands on whoever receives the alert: a clinician's attention, an engineer's wasted inspection, or a member of the public who is searched or refused. Where that cost falls outside the organization, the measurement is most often the one nobody took.
Propositions
A proposition is a claim that holds across more than one case, published with stable identifiers at Propositions. Each one names the cases it rests on; case pages derive their citation from that rather than restating it, so the two cannot drift apart.
Propositions carry their own strength labels — recurrent, emerging, conjecture — deliberately separate from the claim labels above, so that a count of cases is never mistaken for a grade of evidence. A proposition can never outrank the cases beneath it, and the content validator enforces that: a label whose supporting cases are too few, too narrow, or too weakly graded fails the build rather than reaching the page.
Where two cases point in different directions, the disagreement is published as a tension rather than dropped, and the validator warns when a published case is cited by neither.
Causal restraint
Every case is labeled descriptive, associational, plausibly causal, or causal. Most operational reporting is descriptive rather than causal. The label prevents an observed result from quietly becoming a claim that AI produced it.
Review and correction
Records move from lead to research, review, publication, and—when necessary—archive. Published cases show verification and review dates. Material changes are recorded in revision notes, while Git preserves prior versions.