A library of real AI failures. We find the patterns so you can build better governance.
See the casesI came to this work from two directions at once. I studied criminology and volunteered with victims of crime. What victims want is justice, and justice is hard to get without a clear, verifiable record of what actually happened.
I've also spent more than twenty years inside enterprise technology, in the institutions where "we don't know what happened" is not an acceptable answer. The trigger was the pattern. Every week, another story. Air Canada's chatbot. AI lawyers citing fake cases. A coding agent that wiped a production database. Support bots talking in the voice of the previous bot.
Different industries, the same failure: nobody could produce a clean record of what the AI had actually done. So I started writing the incident reports. And the code. Both are free. Both will always be.
Most cannot answer any of them. If the answer to any of them is "we're not sure," that is where Headlights starts.
Every commercial aircraft carries a flight recorder. Almost no AI agent does.
And the record that would have caught each one. New entries arrive every week.
The airline argued the bot was a separate legal entity. The tribunal disagreed. The case turned on a single customer screenshot. The airline could not produce its own record of what its bot had said. Now cited in every legal analysis of AI agent liability.
Read the note →A new agentic chatbot collided with five-year-old scripts the new system had been built on top of. The persona spoke in the voice of its predecessor: drift that nobody documented, and nobody could explain after the fact.
Read the note →
Fifty years ago we solved this problem for humans. Hiring paperwork, reporting lines, conduct policies, performance reviews, personnel files. That wasn't bureaucracy. It was how anyone, a corner shop or a global bank, could prove what their staff actually did.
AI agents are the new workforce. Faster, cheaper, scaling to anyone with an API key, no conscience built in. The paperwork that worked for humans has to be rebuilt for agents.
That's not just an enterprise problem. It's an everyone-shipping-AI problem.
The opposite of a correct statement is a false statement. But the opposite of a profound truth may well be another profound truth.
Eight governance modules in two layers, mapped to a public IETF draft for AI agent audit trails. Apache 2.0, live now on GitHub. Anyone can read every line. Anyone can verify the signatures. No vendor lock-in, no proprietary auditor in the loop.
Read about the code →
No form. No funnel. No auto-responder. No product to sell, no pricing page, no waitlist.
ellie@useheadlights.com →