The map-first principle

Every dataset begins as a pre-registered map.yaml before any item is generated.

Prove first, word second

Code constructs a formal state and proves the key before any language model renders prose. For a puzzle, the builder samples a hidden solution, adds clues that are true in it until the answer is fixed, and a solver checks the result. A model is then asked only to word the item, and blind readers check that the wording still says what the state says.

A kind of item that cannot be proved by construction is marked judged. It enters only through a multi-family unanimity gate, and it is never treated as proved.

Pre-registered cells and capacity

map.yaml declares the dataset’s axes, every intended cell with its target and the generator’s capacity for it, axis share floors and ceilings, variety bounds, and the acceptance thresholds, all before any item exists. A cell the generator cannot yet build stays in the map with capacity: 0 and an explicit reason. A coverage gap is therefore visible, and the map is never shrunk to hide it.

One fingerprint

A project binds project.json, constitution.json, terminology.json, lanes.json (derived from the map’s kind axis) and rubric.json together with map.yaml and canon/. A single deterministic SHA-256 says exactly which configuration a build used.

Where the gates come from

The item checks are the Stage 1 data gates in aspire-si-gates, and the puzzle generator and its solver come from role-set. rdlab composes them into frozen parts with per-item provenance.