Measure the promise: Tabula Rasa
Students prove they understand a subject by teaching a blank AI, fact by fact and rule by rule, until it can pass an interview graded against the teacher's rubric.
It flips the usual worry about AI in school. Instead of a student using AI to skip the work, the AI starts empty and the student has to do the work to fill it. You can't memorize your way through explaining a subject to something that knows nothing.
Measure the promise instead of trusting a prompt. The whole product rests on one promise: the bot knows only what the student taught it. Language models want to be helpful and answer from everything they already know. For months that promise was a paragraph in a prompt and nothing tested it. Now there's an eval: taught questions it must answer, untaught ones it must refuse, and attempts to talk it out of its rules. A model change ships only after that suite passes.
What I tell the agents working on it:
- A stub proves our side only. "A stub is not an LLM."
- An unreadable grade is not a failing score. When grading a student's work, returning no result is better than recording a wrong one.
- Server Actions are public endpoints, so a test calls every one as the wrong person and checks it refuses and writes nothing.
- The framework traps that cost us once are written into the agents' instructions, so the next agent doesn't repeat them.
- Changes ship through small pull requests backed by the live blank-slate evaluation suite.
- Every server action is covered by an automated authorization sweep that calls it as the wrong user and verifies it writes nothing.