Agent incident window › Methodology

How this is collected, and what it cannot tell you

Every limit we know about, stated before you find it.

This AI-agent incident collection methodology names the archive's known gaps. The methodology keeps those gaps visible because an AI-agent incident search cannot measure what its queries never surfaced. This collection therefore treats every known source limit as part of the result.

The seven queries

Four CourtListener RECAP full-text searches, two Bing News RSS queries, one Hacker News (Algolia) date-sorted search. Each is read once a day and the response is sealed with its own SHA-256 digest. Google News RSS is not read: its robots.txt disallows it to us.

Coverage

36 seals across 47 calendar days, 2026-07-10 to 2026-08-25. 240 recorded reads across 7 queries, of which 0 did not return cleanly, and 141 distinct payload digests.

What we deliberately do not publish

No headline, no article summary, no forum post text, no case caption, no docket number and no source URL appears anywhere on these pages. Two separate reasons, and both are load-bearing. First, republication rights over that text were never reviewed, and an unreviewed right is a reason to withhold, not to footnote. Second, the sealed docket snippets carry attorney signature blocks and, in one case, a private individual's personal email address. The exclusion is enforced in the database reader itself — a fixed column allowlist — rather than in any page template, so no page anyone adds later can reach those fields.

The disclosure gate, counted on this request

Counted on this request, over the columns we publish and the columns we exclude — both, because a census that only sweeps the published columns cannot tell “the gate is holding” apart from “the pattern is broken”.

The second number is the control. It is not a leak: those columns are never selected by anything that renders a page. If it ever reads zero as well, the census is broken and the first number means nothing.

What this archive cannot tell you

It cannot tell you how many AI-agent incidents happened. It is what seven keyword searches surfaced, and a real incident that none of them matched is not here. It cannot tell you that an item was removed from anywhere: items leave because a recency ranking pushed them out. It cannot tell you when something was published — the three sources date items in three different vocabularies and we do not order anything by them. And it is not a severity assessment: nothing on these pages ranks, scores or characterises any incident.

The most recent reads, with their digests

One row per query per day. The digest is the first characters of the SHA-256 of the response body, so a reader can see for themselves which days returned an identical file.