PRISM · Release v1
Methods
Eligibility
A launch is included when it is an AI product with its own site, screened in two stages: first its title and URL, then the text of its own page.
Sampling frame and window
The sampling frame is Show HN posts in the sampling window, 2025-09-01 to 2026-08-31, drawn in a seeded order.
Collection
Evidence is collected from each launch's own site and own repository only. The paid search collector was replaced by a free own-site collector for all bulk coding (addendum A8), so every launch was collected the same way.
Coding states and the self-claim rule
A check counts as publicly evidenced only when the launch's own page quotes it and links an artefact; a quote with no linked artefact is coded as claimed. Every check on every launch is in one of four states:
pass: publicly evidenced (artefact linked)claimed: claimed, no artefact linkedunknown: no public evidence foundfail: own page shows the check is not met
AI use
Every role below is an agent, named by its model id and host. Agreement figures are agreement with an independent labeller.
| Role | Model id | Host | Run settings |
|---|---|---|---|
| screener | nvidia/Nemotron-3_5-Lightning | Nebius Token Factory | prompt inclusion rules (protocol section 2) · codebook v1.0 · seed ae0b3a26c3b17672 · token cap 20,000 |
| coder | deepseek-ai/DeepSeek-V4-Flash-0731 | Nebius Token Factory | prompt codebook prompt, guidance c1 on one check · codebook v1.0 · seed none (temperature 0) · token cap 20,000 |
| second coder | nemotron-3-super | Ollama Cloud | prompt codebook prompt · codebook v1.0 · seed 1a8a5e2c0f6afdcb · token cap 20,000 |
| independent labeller | glm-5.3-flash | Ollama Cloud | prompt codebook prompt · codebook v1.0 · seed d1481ad50ed309ea · token cap 20,000 |
| confirmer | nemotron-3-ultra | Ollama Cloud | prompt confirmation prompt · codebook v1.0 · seed none (temperature 0) · token cap 20,000 |
| screening auditor | kimi-k3 | Ollama Cloud | prompt screening audit prompt · codebook v1.0 · seed fa23e4fccbc39ce5 · token cap 20,000 |
| recall gatherer | kimi-k3 | Ollama Cloud | prompt recall gathering prompt · codebook v1.0 · seed cc45c03dd0094b92 · token cap 20,000 |
Confirmation
Every positive is confirmed by a confirmer that did not code it before it enters any figure. A rejected positive counts as not evidenced and stays in the denominator; a downgraded pass counts as claimed.
The deterministic gate
The deterministic gate, not a model, decides each decision from the confirmed check states.
Disclosure control
k = 10. Every figure is a count of launches under disclosure control: a count, its complement or its group size under 10 is suppressed, and an exact audit requires at least 5 possible values for every suppressed count. The complement clause applies: a count whose complement is under 10 is suppressed even when the count itself is not.
Exception (addendum A14): 4 process figures describe the instrument, not any launch, and are published outside the k = 10 rule: screening error of the audited exclusions; supported and raw collection recall; the count of labels edited by a person; the total share of proposed positives changed at confirmation. Per-check precision and per-check changes at confirmation stay under the rule.
Hypotheses and tests
H1 and H3 use one-sided exact binomial tests and H2 a permutation test with Cramér's V. Holm's correction holds the family-wise level at 0.05, and a hypothesis is supported only when it also holds with the measured collection recall as a bound. Each headline figure is repeated excluding the 0 launches whose site could not be read (protocol section 3).
Release ledger
Release v1 is recorded in disclosure ledger v1: later releases are audited against every figure already published, so no hidden count can be worked out by comparing releases.
The codebook and its crosswalk to the standards are in Appendix A of the review.