ReIN Bot
Guess the anime from its opening. 2–8 players, 20-second clips, no accounts — running entirely on free tiers.
A 20-second round, replayed. Guesses are graded server-side; the client never holds the answer.
Everything to the right of the line is unreadable by the client. That placement is the entire security model.
normalise both sides
case · punctuation · long vowels · "season 2" ≡ "2nd season"
then match in tiers
1 exact
2 near bounded Levenshtein
3 season-lenient
4 prefix
all correct tiers score the same;
the tier is only reveal flavour.Biased deliberately toward false negatives — wrongly rejecting a right answer is a smaller sin than accepting a wrong one.
Clips are curated in CI, not at request time
A GitHub Actions workflow reads a manifest of credit-free openings, pulls each from AnimeThemes, cuts 20 seconds, re-encodes to 480p VP9 + Opus, strips every scrap of container metadata, names the file by a deterministic UUID, and uploads it.
It runs on a free Ubuntu runner because the source material is ~5.5 GB and VP9 encoding is slow. None of it touches a developer machine.
The audit
A supervisory audit of the finished schema found and closed two independent paths by which a client could read the answer, four correctness bugs, and three places where the curation pipeline’s spoiler filter failed open.
The lesson written up in PROGRESS.md is the uncomfortable one: execution testing passed because the tests confirmed what the schema’s comments claimed, and the comments had drifted.
Every one verified credit-free, unsubbed, non-spoiler and SFW from the source API rather than assumed.
Rein-Bot
Pick a universe, pick a deck, call the strongest stat on your top card. Two to four players, live.
continue