Developer Journal

Developer Journal

July 2021

Started the project by learning how to draw with the PySide6 library and building some classes to hold crossword puzzle data.

August 2021

Add tools for editing the grid and clues. Publish first puzzle on Stack Exchange, and publish first release of the software.

Add suits for larger grids.

May 2022

Start working on a laser-cut version. The first attempt in wood didn’t turn out well.

June 2022

Second laser-cut attempt with Game Crafter. Proof of concept is a set of tetrominoes for the Donimoes project.

Also start to add warnings when editing a grid and blocks.

Aug 2022

After a successful proof of concept, start working on packing a set of puzzles on both sides of a chip board punchout. Also display hints on which block shapes are needed to balance a set of puzzles.

Sep 2022

Avoid problems with cutter drift, and purchase the first full prototype of a puzzle set.

Oct 2022

Start working on a pair of puzzles in a frame, with jigsaw puzzle tabs to join the pieces firmly. Finding a complementary set of blocks that don’t have identical solutions is hard! Experiment with evolutionary search.

First full prototype arrives, and it worked. Some nicks didn’t hold, and the plain white backgrounds are boring. Try textured backgrounds and remove duplicate cut lines.

Experimenting with changes to evolutionary search. The totally random mutation averages 199 +/- 121 epochs over 33 trials to fill the back grid with no restrictions. Mutating by removing the blocks closest to one of the gaps averages 224 +/- 119 epochs over 204 trials. Bizarrely, adding pairing made it worse! Not only did it increase to 481 +/- 216 epochs over 29 trials, but each epoch was noticeably slower.

Nov 2022

Switch from background wood texture to shaded pattern. Distribute background colours around the colour wheel. Expand to two-part frame for 11x11 pair.

Dec 2022

Switch from HSV to JCh colour model, now lightness looks consistent across different hues. Dark suits are still hard to see against the background, though. First prototype of a puzzle pair arrives. Trim the outside of the 9x9 pair’s frame, since the background doesn’t bleed all the way to the edge.

Start talking to a manufacturer about bigger production runs.

Jan 2023

First double frame prototype arrives, as well as JCh colour model. Dark suits are better, but still hard to see. Use a different font for each puzzle in a set. Rewrite the musical set to avoid one blank piece.

Jan 2024

The prototypes are working OK, but a manufacturer said that the piece tabs are too small to punch out. Packing without rotation would let me use single tabs instead of the rotationally symmetrical double tabs, so they could be twice as big. Start working on packing without rotation, which is much harder.

Feb 2024

Try calculating all available slots for each shape before packing begins.

Dec 2024

Try packing both sides at the same time.

Feb 2025

Add a two-sided evolutionary block packer.

May 2025

Try building a set that’s printed on one side, with blank dry-erase surface on the back. Then you could copy the letters for any puzzle onto those pieces.

Sep 2025

First prototype of dry-erase pieces arrives, but copying is way too fiddly. You have to find the right shapes, copy the letters and the numbers, and then solve. It takes longer to set up than to solve.

The single tabs are nice, though.

Weirdly, the printed side has over 5mm of drift this time, and is unusable. I don’t think any of the other prototypes have had more than 1mm of drift. Very disappointing!

To try and get the packing without rotation to work, I discovered Knuth’s algorithm X for exact cover problems, which the packing seems to be. Luckily, our library provides access to O’Reilly’s copy of The Art of Computer Programming. On first read through, it seems promising, and I found the xcover library that seems to have implemented it in Python with Numba just-in-time compilation.