Bizarre Patent Game (Poppy): guess what a real patent is for

Poppy is a browser game built on real, bizarre US patents: the comb-over, the method of swinging on a swing, the machine that kicks its user. It has two modes.

• Guess the patent. One drawing appears. Type what the patent is for, or ask for the next clue: a second drawing, then a passage from the patent with the giveaway words blanked out. The first clue is worth up to 100 points, the second up to 70, the third up to 40. The reveal shows the title, number, year, owner, every drawing and a link to Google Patents.
• Real or fake. A real patent sits beside an invented one, each with a drawing and a passage. Pick the real one for 50 points.

Try it: open index.html from any web server (python3 -m http.server). No account; the score lasts for the browser session.

How guesses are scored

Every patent has a plain-English description of 8 to 12 words ("hiding a bald spot by combing long side hair over it"). A small sentence-embedding model (all-MiniLM-L12-v2, run in the browser by transformers.js) turns the guess and the description into vectors. Their cosine similarity is rescaled to 0 to 100 (0.30 and below is 0; 0.72 and above is 100) and capped by the clue level. Scoring is by meaning, so "kicks you in the butt" does well against "a machine that kicks the user in the buttocks". The official titles are not used for scoring: they vary too much in length and wording.

The model (about 35 MB) is downloaded once and cached. If it cannot load, the game falls back to counting shared words.

The data

data/patents.json holds one entry per patent: number, official title, description, issue year, owner (assignee, or first inventor when unassigned), figures in clue order, excerpt, Google Patents link, and whether it is real or invented. Invented entries carry only a title, a figure and an excerpt; their drawings are borrowed from ordinary real patents, credited at the reveal.

The drawings in img/ are the patents' own sheets. Whole scanned sheets print the title across the top, which would give the answer away, so tools/build_images.py erases the header lines and drops pages of text.

To rebuild or extend the set:

python3 tools/fetch_patents.py numbers.txt          # title, dates, owner, text and sheets from Google Patents
python3 tools/build_images.py work/img img numbers.txt > work/manifest.json
python3 tools/build_data.py work .                  # needs work/llm/out_*.json and work/fakes.json

The descriptions, excerpts and invented patents were written with Claude and checked by script (word counts, required fields, files present).

Name

Poppy: the red wildflower that turns up where nobody planted it.

MIT License for the code. Patent text and drawings are from the public record.