It learns your system from cases you’ve already solved.
Then it proves the gain on cases it was kept away from, and ships only if Claude gets more right.
From your history to a release you can trust.
The same five steps every time. Nothing ships on a hunch.
- 01Read-only
Connect safely
Day Theory runs on your machine with a read-only role. It won’t start unless the role can only read and everything it can reach is on your allow-list.
- 02Blind split
Split your history
Your solved cases are split blind into a learning half and a test half. Drafting never sees the test half.
- 03Checked rules
Draft and check
An analyst agent studies the learning half and writes a short set of rules about your system. Every rule is checked against your data.
- 04Paired test
Prove it
Claude answers the test half with and without the knowledge. A paired statistical test compares the two.
- 05Release gate
Ship only if it helps
It ships only if Claude gets significantly more right and no group of questions gets much worse.
What it needs today: cases your team has already solved. Drafting from the data alone isn’t built yet.