The Corpus, At A Glance.
Measured post-mortems of agentic engineering failures — every one paired with the control that prevents it. This is the overview; every number here drills into reports or tags for the case-by-case detail.
When an AI agent gets corrected mid-session, the correction usually dies with the session — the next person, on the next machine, hits the same wall. There’s no shared record of the everyday engineering failures agentic AI actually produces.
This is that record: a public registry of measured engagements — what was attempted, what it cost, what was wasted, and for every failure, the specific control that would have prevented it. Figures are traced to a source and reconciled, never self-reported.
It is deliberately not a vendor scorecard, not an AI-harms incident database, not a benchmark, and not marketing. Just what actually happened, measured, with the fix attached.
A failure without a control is an anecdote.
- Dashboard
- The aggregate view — you’re on it now.
- Reports
- Each engagement, told in full, one per page.
- Tags
- Every failure, addressable by what kind of mistake it was.
- Feed
- Newest reports first.
Detection route and silence rate aren’t shown yet — the source reports don’t tag those per-failure (only 1 of 14 does today on report 001). They’ll appear once that’s retrofitted.