zot whetstone

A software factory that hones one game to perfection.

https://openzot.github.io/whetstone/

The arcade ships a new game every shift; the machinery a new machine. The whetstone ships no new thing at all. It holds a single game - and every shift, zot plays it, reads the log of every pass made before, finds the dullest facet, and makes exactly one improvement: sharper feel, a fairer difficulty curve, a sound where there was silence, a cleaner game loop. Then it proves the pass by playing again, and writes one entry in the ledger. One blade, many passes.

Some notes

• Nobody reviews the passes before they ship. They are model output, published as-is; expect the occasional pass that only polishes the flat of the blade.
• The standing order is orders/hone.yaml - the whole specification of what one pass across the stone must do. It never changes; what changes is the game, and site/ledger.json, the honing log every shift reads first and appends to last.
• Every version is kept: a pass copies the latest version to the next numbered folder under site/versions/ and hones the copy, so the whole history of the edge stays playable - version 0 is the game as it came off the casting bench, version n is the game after pass n, and the log page lines each version up with the entry that explains what changed and why, next to a recorded mid-play screenshot of that version (preview.png, taken by the probe), so the whole evolution is visible at a glance.
• AGENTS.md is what zot reads before every shift: the facets it may hone, the rule that it must play before it may touch, and the one it must never break - the blade stays sharp; the game is playable at the end of every shift, or the shift does not land.
• The gate is in two halves: scripts/check.sh holds the shape (static - one folder per pass, three files each plus optional assets, a valid ledger), and scripts/probe.sh plays the latest version in a headless browser - it must load clean, start, run and reset - and takes the screenshots the model is required to look at.
• Setup, workflows, layout and tuning are in OPERATING.md. The short version: fork this repository, add an OPENROUTER_API_KEY secret, and run the shift workflow once.
• Every shift's session - the whole conversation of playing, judging and honing - is published as a row in the openzot/whetstone dataset on Hugging Face, for fine-tuning, training or evaluation; how rows are produced is in OPERATING.md.
• Why the product is fixed and only the facet varies - and why that is still a factory and not a workshop - is in PHILOSOPHY.md.