The key comes from code, never from a judge.
rdlab plans a dataset as a pre-registered map, builds items whose answer keys are constructed and proved by code, asks a local model only to put them into words, and records the provenance of every item.
Install
pip install reasoning-dataset-lab
# Python 3.13 or newer.
Check a map
rdlab map check examples/t10-pilot-a.map.yaml
# validates the cells, capacities and share rules,
# then prints the map's canonical SHA-256.
# no model, no network.
Fingerprint
rdlab project fingerprint . --map examples/t10-pilot-a.map.yaml
# one hash for the five project files,
# the map and canon/.
How it keeps the key honest
A key that a model chose is an opinion. A key that code built and proved is a fact about the item.
Prove first, word second
Code builds the item from a hidden formal state and proves the key. A model is asked only for the wording, and blind readers from different families check it.
Pre-registered cells
map.yaml declares every cell, its target and the generator's capacity before any item exists. A cell the generator cannot build stays in the map with capacity 0 and a reason.
One fingerprint
Five project files, the map and the canon hash to one SHA-256, so a build says exactly what it was built from.
Certified value sets
The values a claim is swapped within are certified before use. In the v3 run 44 of 50 sets certified, against 26 before repair. It is evidence, not proof, and the page says so.
Local and bounded
Models run through Ollama on loopback. Anything else is refused. rdlab never pulls a model and never calls a cloud endpoint.
Not training data yet
Nothing rdlab writes is training data until a person has checked it.
Use it
Install
pip install reasoning-dataset-labCheck a map, then fingerprint the project
rdlab map check examples/t10-pilot-a.map.yaml
rdlab project fingerprint . --map examples/t10-pilot-a.map.yaml