PRISM · Release v1

What AI launches publicly evidence

Release v1 · 2026-10-25 · 300 included launches

Hypotheses

Launches that claim at least one check without linking an artefact

32%96 of 300 · 95% CI 27–37%

H1 (a majority of launches claim at least one check without linking an artefact, so they do not publicly evidence it) is not supported: 96 of 300 launches (32%), one-sided exact binomial, Holm-adjusted p = 1.000; with the recall bound, 85 of 300, Holm-adjusted p = 1.000.

Evidence by discovery source

Reported descriptively: at least one source has fewer than 30 included launches, so the pre-registered test was not run. The Holm correction covers the remaining hypotheses.

Launches that publicly evidence at least one rely-decision check

71%212 of 300 · 95% CI 65–76%

H3 (fewer than half of launches publicly evidence at least one rely-decision check) is not supported: at least 212 of 300 launches (71%) publicly evidence one, one-sided exact binomial, Holm-adjusted p = 1.000; with the recall bound, 238 of 300, Holm-adjusted p = 1.000.

How we know

Launches counted: N = 300 included launches. Launches with no evidence found stay in the denominator.
Agreement: Agreement with an independent labeller: AC1 0.72–0.89 across the shown checks (not estimable for 20); per check on the Validation page. Agreement not estimable for 3 shown checks (fewer than ten gold passes); those figures are descriptive. Labeller: glm-5.3-flash on Ollama Cloud.
Changed at confirmation: 39.4% of proposed positives changed at confirmation (26% rejected, 14% downgraded to claimed). Confirmer: nemotron-3-ultra on Ollama Cloud.
Who checked: Agent roles, no human edits. Share human-edited: 0%.
Sensitivity: Excluding the 0 launches whose site could not be read: Launches that claim at least one check without linking an artefact: 96 of 300 (32%, 95% CI 27–37%); Launches that publicly evidence at least one rely-decision check: 212 of 300 (71%, 95% CI 65–76%).
Release: Release v1, 2026-10-25. Data: public-data/release.json.

Download the review (PDF)

Reading the figures

Every check on every launch is in one of four states: “publicly evidenced (artefact linked)”, “claimed, no artefact linked”, “no public evidence found”, or “own page shows the check is not met”. Launches with no public evidence found stay in every share’s denominator. Not publicly evidenced does not mean not done.