AgentGlass lab notes

Measured, not demoed.

Launch-day claims about AI models arrive as demos and hand-picked samples. These write-ups test the claims instead: the prediction is committed before the first result, the method is stated, and the misses are reported as plainly as the hits. Copy a rule that held up, not a demo that looked good.

01 · PredictionCommitted firstEvery claim is written down before the test runs.
02 · MethodStated in fullPrompt, settings, reader and scorer, so you can rerun it.
03 · NumbersTraced to filesEach number in a post comes from a data file.
04 · The missReported plainlyA miss gets the same box and the same weight as a hit.
Write-ups

Latest

Model tests, FactFrame notes and build notes, newest first. Each one opens with what was predicted and what happened.