A safety net for
AI that makes small mistakes.
TDI watches an AI model while it works, catches the slips that matter, and has the model correct them quietly — before they reach you.
Good AI still slips — and the slips are the expensive part.
Modern AI models are impressive, but they make a particular kind of mistake: small, confident and easy to miss. They forget something you said earlier. They write code with a typo in it. They call a function that sounds real and isn't. Each one reads perfectly well, which is why people only find out later, when it has already cost them time or money.
Making the model bigger does not remove this. So we asked a different question: can a small, separate layer watch the model and catch these slips as they happen?
It sits beside the model, not inside it.
TDI works alongside any open-source language model without changing the model itself. When it spots a slip, it sends the model a short reminder you never see, and the model corrects its own answer. You get the corrected version. There is no warning to click through, and nothing is held up while it works.
- Works with the model you have. It is not tied to one model or one vendor.
- Invisible when it works. Corrections happen behind the scenes; the model is nudged, not overridden.
- Quiet by design. A checker that cries wolf is worse than none, so we measure wrong corrections as carefully as missed ones.
Three slips, with and without TDI.
Strong where we've finished testing.
Results for the first two kinds of slip, from our own tests on data we held back during development. Internal results — not yet independently verified.
of answers that broke a stated instruction were caught
of the AI's own slips were fixed — against about 13% when left alone
of corrections were wrong (0.32%), meaning TDI rarely “fixes” something that was fine
Measured on broken instructions and on broken code. Both sets were tested on examples TDI had not seen while it was being built.
Anywhere a small AI slip has a real cost.
These are the fields where we think TDI is most useful. We have not yet tested it inside any of them — they are where we're aiming, not where it is proven.
Customer support
Assistants that must keep to a refund policy, a promised price or a customer's stated constraints across a long conversation.
Software development
Coding assistants that write real code against real libraries, where one invented function or broken file wastes an hour.
Finance and operations
Workflows where a budget, a limit or an approval rule given at the start must still hold at the end of a long task.
Legal and compliance drafting
Drafts that must keep to the terms and exclusions a reviewer set, instead of quietly drifting from them.
Education
Tutors that should stay consistent with what a student has already been taught, and with the rules of the exercise.
Long-running AI agents
Agents that work for many steps on their own, where a small early slip compounds into a wrong final result.
What's done, and what isn't.
Two of the five kinds of slip we're targeting — broken instructions and broken code — have reached our bar in testing. Made-up functions are close but not there: TDI currently catches about 92% of them, below the 95% we require, so we are not claiming it. Staying on task and contradicting itself are not built yet, and we haven't run the combined test across all of them together.
TDI is not a product, not available to use, and has not been tested with real users. We'll publish more as results earn it — and say so plainly when they don't.
Working on AI that has to be right?
If a field above is yours, we'd like to hear what a slip costs you. Write to us directly.