Open research material · catalog v1

The data behind the deduction.

512 fixed puzzle records, their stored grades, and measurements of one reproducible solving path. Download the records and inspect how each number is defined.

Download CSV · 512 rows Download JSON + example states

Population and selection

This is the full frozen v1 catalog, with 128 entries per grade. The generator uses seeded completed-grid generation, shuffled clue removal with uniqueness checks, and grade quotas. It keeps at most one puzzle per generated completed grid and prefers the highest available accepted grade. Its conservative clue-mask fingerprint filter is not a uniform sampler over all Sudoku puzzles.

The dataset contains no player records. The computational path completed 512 of 512 entries; 0 stalled and 0 returned invalid. This is a recorded output from the same logical engine, not an independent mathematical certification.

Example selection

Selection rule: first Easy entry in catalog order with a higher-clue entry in another grade, followed by the first such higher-grade entry. This is an illustrative pair, not a representative sample.

Field dictionary

id / puzzle / solution: Stable published ID and 81-digit strings; 0 denotes a blank in puzzle. Import these columns as text in spreadsheets to preserve leading zeroes.
difficulty / hardestTechnique: Stored v1 labels, derived from the highest-ranked technique used in the original grading path.
clues: Number of nonzero digits in the original puzzle.
firstEliminationStep: One-based position of the first productive candidate-elimination step. Empty in CSV / null in JSON when the path contains only placements.
emptyAtFirstElimination: Empty cells immediately before that first elimination; not remaining solving time.
placements: Number of placed digits in the recorded path.
productiveEliminationPasses: Number of successful elimination-rule applications; one pass may delete multiple cell-digit candidates.
explicitCandidateRemovals: Cell-digit candidates removed by elimination techniques. Routine peer updates after a placement are excluded.
traceStatus: The current path computation’s result: solved, stalled or invalid. This is not an independent proof check.
recordedSteps (JSON): Original per-technique counts: singles count placements, other techniques count productive passes. Do not add them and call the result human thinking steps.
examples / hardCase (JSON): Selected steps with complete before/after values, logical candidates, support cells, source/target units and exact actions. Arrays use zero-based cell indexes; diagram labels are one-based r/c.

Reproduce this edition

Use Node.js 22.18 or later. Save the three source snapshots at the paths shown, then run the exporter from their project folder. It reads the catalog and writes only research assets; it does not regenerate or overwrite the frozen catalog.

node --experimental-strip-types scripts/export-research.ts

Version and provenance

Dataset: catalog-v1
Trace model: v1-first-supported-2026-09-06
Catalog SHA-256: 3a075495ccbc8d349ea4ad0f75b3798eab4c37f2bf13fa504adea08f0ef6056a
Grader source SHA-256: ea418c8cb5b38271a2ef1d347a2959bc3cd09e1760c784b9c44e1f6e49cc9655
Exporter source SHA-256: 18fb5b19e95a96d496ef43d68903cca62a37d28657bf66ac37a66287bc274498

Solver configuration

There are no chains, triples, fish or uniqueness-based deductions in this path model. Technique order and within-technique scanning order are fixed. A technique used on one path is not necessarily required on every path. Grade quotas bias the sample; this is not a prevalence estimate for the wider Sudoku population.

The median is the middle value after sorting, averaging the two middle values for an even-sized group. Counts are complete descriptive measurements of this catalog. We do not attach population confidence intervals to this fixed, selected collection.