The replay test

> DRAFT, machine-compiled, owner review pending.

What it asks

For each of the 37 Index incidents with a public trace date, would the dated public archives a watch loop reads (Hacker News, arXiv, the AI Incident Database, GDELT) have held a matching item within 24, 48 and 72 hours of the first public trace?

Result (results file results-2026-10-12.json)

Window	Caught	Rate
24h	19 of 37	51%
48h	26 of 37	70%
72h	29 of 37	78%

On the 13 incidents whose trace date is exact, the rates are 77%, 92% and 92%.

Never surfaced in 72h (8): RIA-001 OpenAI ExploitGym agents breach Hugging Face; RIA-003 OpenAI-linked agents use a German programming wiki (DseWiki) as a private messag; RIA-010 Self-propagating prompt injection ('worm') passes between agents in a controlled; RIA-017 Grok Bot 'personal CFO' allegedly posts founder's bank balances to company Slack; RIA-024 Alibaba-affiliated ROME agent opens reverse SSH tunnel and diverts GPUs to crypt; RIA-025 Replit agent deletes production database during code freeze; RIA-026 Amazon Q Developer VS Code extension shipped with a wiper prompt; RIA-036 AI Village agents send unsolicited email, fabricate claims.

Limits

It measures the archives, not Ria's screening model; a hit means a matching item existed.
The incident set is the development set the source list was tuned against, so the rates are optimistic for unseen incidents.
Trace dates are day-grain and some are approximate.

Full report and method: https://ria.arnao.ai/research/replay-backtest-2026-10-12.html