dedoku 1.1.0

A pure-Python Sudoku solver that uses only human-style logical deduction techniques (no backtracking).

A pure-Python Sudoku solving library that relies exclusively on human-style logical deduction — 20 named technique families, from naked singles to alternating inference chains, and not a single guess.

About the project

Most Sudoku solvers brute-force the board: try a digit, propagate, undo on contradiction. This library takes the opposite stance. Every digit placed and every candidate eliminated is the conclusion of a named, explainable technique, applied exactly the way a strong human solver would reason. The full solving path is recorded step by step, so any solution can be replayed and audited.

• Zero external dependencies — Python 3.10+ standard library only, tests included (unittest).
• Fully typed and documented — modern type hints and Sphinx-style reStructuredText docstrings on every module, class, and method.
• Explainable solving — each Step reports the technique, a human-readable description, the placements, and the eliminations.
• Honestly benchmarked — a reproducible 500-puzzle benchmark against a reference backtracking solver ships with the repo (below).

Solving philosophy

Logic first. Guessing only if you ask for it. By default the solver never backtracks: if the pipeline of logical techniques cannot finish a puzzle, it stops and says so (11 puzzles out of 100 on the benchmark's hardest tier). When you just need the answer, two explicit opt-in modes exist — method="hybrid" completes the remainder by brute force after the techniques stall, method="backtracking" skips logic entirely — and anything brute-forced is recorded as an explicit "Backtracking" step: the solving path never lies about how a cell was filled.

Installation

pip install dedoku

Or from source:

git clone https://github.com/n36l3c7/Dedoku.git
cd Dedoku
pip install -e .

Either way, no dependencies come with it — the package is pure standard library.

Usage

The one-liner:

import dedoku

result = dedoku.solve("530070000600195000098000060800060003400803001"
                      "700020006060000280000419005000080079")
print(result.solved)   # True
print(result.grid)     # pretty-printed solved board
for step in result.steps:
    print(f"[{step.technique}] {step.description}")

# When you just need the answer, even beyond the logical pipeline:
dedoku.solve(puzzle, method="hybrid")         # logic first, brute force finishes
dedoku.solve(puzzle, method="backtracking")   # brute force directly

Or with full control over the board and the pipeline:

from dedoku import Grid, SudokuSolver

puzzle = (
    "530070000"
    "600195000"
    "098000060"
    "800060003"
    "400803001"
    "700020006"
    "060000280"
    "000419005"
    "000080079"
)

grid = Grid.from_string(puzzle)
result = SudokuSolver().solve(grid)

print(result.solved)          # True
print(grid)                   # pretty-printed solved board
print(grid.to_string())       # 81-character string

for step in result.steps:     # the full, explainable solving path
    print(f"[{step.technique}] {step.description}")
print(result.techniques_used) # distinct techniques, in first-use order

Puzzle strings are 81 characters, read left to right, top to bottom: digits 1–9 are givens, 0 or . mark empty cells; whitespace and the decoration characters |, -, + are ignored, so pretty-printed boards parse back.

Need a custom pipeline? Pass any sequence of techniques, tried in order:

from dedoku import SudokuSolver
from dedoku.techniques import HiddenSingle, NakedSingle, XYChain

solver = SudokuSolver(techniques=[NakedSingle(), HiddenSingle(), XYChain()])

The board model is fully navigable if you want to build your own techniques: each Cell knows its row, column, subgrid, candidates, and peers; the Grid exposes all 27 houses via rows, columns, subgrids, units. Full guides and API reference: n36l3c7.github.io/Dedoku.

Command line

Installing the package also installs the dedoku command. Add --explain to watch the whole reasoning, cell by cell:

$ dedoku 530070000600195000098000060800060003400803001700020006060000280000419005000080079 --explain
   1. R5C5 = 5               [Naked Single] R5C5 has 5 as its only candidate
   2. R5C2 = 2               [Naked Single] R5C2 has 2 as its only candidate
   ...
Solved in 51 steps using: Naked Single

Options: puzzle from argument or stdin, --method {logic,hybrid,backtracking} (logic is the default; the other two finish any puzzle by explicit brute force), --multi for puzzles without a guaranteed unique solution, exit codes 0/1/2 (solved / stalled / invalid).

Implemented techniques

The default pipeline applies 28 technique instances from 20 families, always restarting from the simplest after every deduction.

#	Family	Classes	Module
1	Naked Candidates	NakedSingle, NakedPair, NakedTriple, NakedQuad	naked.py
2	Hidden Candidates	HiddenSingle, HiddenPair, HiddenTriple, HiddenQuad	hidden.py
3	Intersection Removal	PointingCandidates, ClaimingCandidates	intersections.py
4	X-Wing	XWing	fish.py
5	Chute Remote Pairs	ChuteRemotePairs	chute.py
6	Simple Colouring	SimpleColouring	colouring.py
7	W-Wing	WWing	wwing.py
8	Y-Wing (XY-Wing)	YWing	wings.py
9	Unique Rectangles	UniqueRectangle (Type 1)	rectangles.py
10	Swordfish	Swordfish	fish.py
11	XYZ-Wing	XYZWing	wings.py
12	BUG (Bivalue Universal Grave)	BivalueUniversalGrave	bug.py
13	Avoidable Rectangles	AvoidableRectangle	rectangles.py
14	Unique Rectangles Type 2	UniqueRectangleType2	rectangles.py
15	Finned Fish	FinnedXWing, FinnedSwordfish	fish.py
16	X-Chain (basic X-Cycles)	XChain	chains.py
17	XY-Chain	XYChain	chains.py
18	3D Medusa	Medusa3D	medusa.py
19	ALS-XZ (Almost Locked Sets)	AlsXz	als.py
20	AIC (Alternating Inference Chains)	AIC	aic.py

Uniqueness-based techniques (9, 12, 13, 14) assume the puzzle has exactly one solution — the standard convention for published Sudokus. Solving a puzzle that might have multiple solutions? Pass assume_unique=False to dedoku.solve() or SudokuSolver() and they are excluded, keeping every deduction sound.

Benchmark

500 unique-solution puzzles (seed 42), 100 for each of five difficulty levels, timed against the lightest classic backtracking solver (bitmask, sequential cells, no heuristics). Same protocol for both solvers: puzzle string in, board out, minimum of 3 runs, parsing included. Python 3.13, Windows 11. Levels are graded by the hardest technique the original 13-family pipeline needs: singles → subsets → intersections → advanced → extreme (beyond that base pipeline).

Solve-time distribution

Level	Solved by library	BT median	BT p95	BT max	Library median	Library p95	Library max
1 · Singles	100/100	18.3 ms	526 ms	1,913 ms	1.3 ms	2.1 ms	2.7 ms
2 · Subsets	100/100	14.0 ms	218 ms	2,314 ms	1.6 ms	2.6 ms	2.8 ms
3 · Intersections	100/100	5.4 ms	134 ms	2,165 ms	3.4 ms	8.8 ms	12.9 ms
4 · Advanced	100/100	11.2 ms	144 ms	353 ms	5.0 ms	12.1 ms	23.2 ms
5 · Extreme	89/100 †	17.0 ms	241 ms	721 ms	60.1 ms	215 ms	495 ms
Overall	489/500	12.0 ms	239 ms	2,314 ms	2.8 ms	104 ms	495 ms

† On the 11 extreme puzzles the library cannot crack, its reported time is the time to exhaust every technique and stop — never a guess.

Key findings:

• The logic library wins on the median at every human-graded level up to Advanced (2.8 ms vs 12.0 ms overall) and is far more predictable: naive backtracking's worst case is 2.3 s when the cell order is unlucky, versus 0.50 s for the library's hardest chain solve.
• Brute-force time is uncorrelated with human difficulty — backtracking is fastest on level 3 and slowest on level 1, while the library's times grow monotonically with the level. The two solvers measure different notions of "hard".
• Level 5 is chain territory: the advanced techniques added in v0.2.0 raised the library's extreme-tier solve rate from 0/100 to 89/100.

Which techniques crack the extreme tier

Reproduce it

python benchmark/generate_puzzles.py   # optional: regenerate the dataset (seeded)
python benchmark/run_benchmark.py      # times both solvers, writes benchmark/results.csv
python benchmark/make_charts.py        # renders the SVG charts into docs/

Architecture

Class	Responsibility
Cell	A single cell: value, candidates, position, given flag, and references to its row, column, and subgrid.
Unit → Row, Column, Subgrid	The 27 houses of nine cells, sharing bookkeeping logic.
Grid	The full 9×9 board: wiring, parsing, serialisation, validity.
Technique / Step	One logical strategy and the immutable record of one applied deduction.
SudokuSolver / SolveResult	The engine that chains techniques and the outcome of a session.

Development

python -m unittest discover -s tests -v   # 100 tests, includes the step oracle
python -m ruff check dedoku tests benchmark
python -m mypy dedoku
python -m coverage run -m unittest discover -s tests && python -m coverage report

CI runs the suite on Linux, Windows, and macOS across Python 3.10–3.14, with ruff, mypy, and a 95% coverage gate. Beyond unit tests, an oracle check verifies that every placement matches an independent backtracking solution and that no elimination ever removes a true digit — the library's core contract, tested directly.

Soundness validation

python benchmark/validate.py --count 100000 --seed 100 --jobs 8

The 1.0.0 release run: 100,000 generated puzzles, every single deduction verified — 98,687 solved by pure logic (98.7%), 1,313 stalled (extreme chain territory), 0 unsound steps.

Versioning and stability

The project follows Semantic Versioning and is stable as of 1.0.0. The public API is everything importable from dedoku and dedoku.techniques and documented in this README; underscore-prefixed names are internal. Breaking changes only happen in major releases, preceded by a deprecation period, and every change is listed in the changelog. Note that SudokuSolver.solve(grid) mutates the grid it receives — use dedoku.solve(puzzle) if you prefer a fresh board per call.

License

Released under the MIT License.