Parity scoring
One number per clone, then the layers. Sealed tests count. Nothing is rounded up.
What the number means
Parity is the pass rate of the conformance suite, visible and sealed together, broken down by layer. The suite is written from the approved spec before any code exists, so the builders are measured against a yardstick they did not choose and cannot edit.
Sealed tests
About 30% of the tests never leave the room. Builders work only from the spec and the visible ~70%. The parity harness is the only thing that runs the sealed set, in a scratch checkout, and sealed results leave it as counts per feature: no test names, no assertion text, no output. Anything more would leak the holdout into the next build round.
A feature that passes its visible tests far more often than its sealed ones is flagged as overfitting, on the run page and in the report.
What stops a run
| Check | What it catches |
|---|---|
| Hash check | Tests or spec changed since they were published. The hashes are recorded at publish and checked on every merge. |
| Stub scan | Code that pretends to implement a feature so a test passes. |
| The harness decides | An agent's claim that its tests pass is ignored. Only the harness's own run counts. |
Scoreboard
| Behaves like | Windows 98-era Paint |
|---|---|
| Stack | Rust + egui, native + wasm |
| Parity | {{BRUSHLING_PARITY}}% |
| UI | {{UI}}% |
| Behaviour | {{BEH}}% |
| File formats | {{FMT}}% |
| Sealed pass | {{SEALED_PASS}}/{{SEALED_TOTAL}} |
| Tests | {{TESTS}} |
| Agent-hours | {{HOURS}} |
| Cost | ${{COST}} |
| Demo | Try it in your browser |
Scores are floored to one decimal. A failing layer is never dropped from the table.
What parity can't see
Parity measures the spec, not the original. A behaviour nobody wrote down is not tested, so a missing spec line is a gap the score cannot see. That is why the spec is published next to each clone: you can check what was measured.
Add your clone
Ran Remakecraft on something? Open a pull request adding a row to the README, with a link to the clone, its parity.json and its wall.log, so anyone can check the score and the wall.