Including what broke. A changelog that only announces wins is marketing, and nobody believes it twice.
The homepage runs a real skill on your machine: it fetches the code, builds a Word document from your values, opens that document back up and checks it in front of you. Nothing is uploaded and nothing runs on our servers, which is only possible because a skill is kilobytes.
You can hold a name, publish under it, follow someone, and post paid work. We take a listing fee and never hold the bounty itself: the person who wanted the work pays the person who did it, directly.
A controlled list of twenty three tasks, and search filters on it. We measured why: ranking by description is flat at a quarter correct from fifty skills to a thousand, while matching on a typed field does not decay at all.
The new re-verification runner caught it on its first run. The forge proves four blocks of code as separate modules but shipped them merged into one file, so a later import silently overwrote an earlier one. The gate had been testing a different arrangement than the file we ship. Fixed, and all nine skills pass again.
Every skill is run again on freshly generated cases and its page says when that last happened. A signature proves the code has not changed; re-running the check proves it still works, on today's libraries.
We described thirteen kinds of job and measured a cheap model on each. It already handled nine of them, so nothing was built. One skill was built, measured, and thrown away because its gain sat inside the confidence interval. The publisher enforces that automatically now.
With a skill wrapped as a single call, a model small enough to run on a laptop scored sixteen out of sixteen on fillable PDF forms and on combo chart decks, with no API and no cost.
