“Tests passed” now means exit 0
Our own tool graded a test run by reading the command, not the result. A cargo test that died on a compile error scored the same as one that went green. Here is what the fix moved, and what it did not.
Writing
Notes from measuring our own agents.
Everything here comes from one index on one laptop, and every number has a spec or a pull request behind it. Where a measurement is narrow, the post says how narrow rather than rounding it into a claim.
What survives compaction
One session compacted eight times kept about a third of one percent per round. We diffed what went in against what came out — 402 decision-shaped sentences in, 28 out, and reasons survived at 1.7%.
The repo predicts success better than the model
Across one index, the spread between the best and worst repository was 75 points. Between the best and worst model, 32. That is the opposite of what a model leaderboard would suggest — and it is one machine's data.
Point it at your own machine.
npx -y agentworth scan