Post-Incident Analysis of AI Failure
Research papers, technical notes, and public-facing publications connected to the work behind the forensic governance frameworks and the broader AI risk literature.
The Reserve Bank's model risk draft is the right framework written in the wrong tense. Here is the feedback I submitted, what the draft already gets right, and the one word that decides whether any of it works.
26 July 2026
The Reserve Bank's data governance draft asks for lineage, metadata and disposal records at every step of a data lifecycle. The one addition that decides whether any of it works is a single word the draft already used once. Here is the feedback I submitted, what the draft gets right, and why it travels far beyond banking.
18 August 2026
We classified all 1,560 AI Incident Database entries. 62.6 percent are normative failures, systems working exactly as designed, resolving two legitimate goals wrongly. Attacks are under four percent. The majority class of AI harm is not untestable. It has never been specified. Developers can fix it now, at the application layer, without waiting for anyone.
2026
When the record enters the objective, the record is captured, and every downstream use of it, audit, oversight, evaluation, compliance, inherits the corruption. The impossibility corollary says the divergence lands exactly where truth is expensive. The unscored-action corollary says the alarm nobody scores is the alarm nobody raises. And the awareness condition says all of it runs on what the agent believes about being watched, which is why better logging alone cannot fix it.
2026
