← Projects

What Year Is It? shows a still from a film and asks for its year. You don’t type a year; you narrow a range in steps, and each correct step banks a point. A wrong step ends the round. Play at wyit.film.

Status
live at wyit.film
Monthly cost
$0
Infra
static CDN, 2 GETs per round
Payload
~198 KB gzip cold first round
Performance
Lighthouse 97 perf
Corpus
4,028 films, 1915–2025

Overlapping windows

The first version used hard buckets: a century, then a decade, then a year. Being off by one was too unforgiving. Now each step offers up to four windows over the current range, and adjacent windows overlap, so a film near a boundary falls in two of them and either is a correct pick.

1900–19491925–19751950–20001976–20251900195020002025(a)
1900–20251950–20001961–19771965–19711966–19681967(b)
Figure 1 Narrowing toward a 1967 film, as the engine computes it. (a) The first step offers four windows over 1900–2025; adjacent windows overlap, so a film from 1950 or 1975 sits in two of them. (b) Each pick becomes the next parent range until the last step offers single years.

The steps come from one plan. Three steps come first; after them an adaptive step keeps subdividing until the range fits in five years, and the last step offers single years, at most six:

initial: [1900, 2025],
steps: [
  { bins: 5, windowBins: 2, strideBins: 1, windows: 4 },
  { bins: 9, windowBins: 3, strideBins: 2, windows: 4 },
  { bins: 9, windowBins: 3, strideBins: 2, windows: 4 },
],
exactWidth: 5, maxSteps: 5,
adaptive: { bins: 9, windowBins: 3, strideBins: 2, windows: 4 },
maxSingles: 6,

A step cuts the parent range into bins equal pieces and makes each window windowBins pieces wide, starting a new window every strideBins pieces.

The input is discrete and buttons would work. The drag pill is there for a physical sense of snapping into one of the options while still implying a sliding, overlapping range.

A slot number is an address

Offline, a seeded shuffle assigns the corpus of FilmGrab stills to slots 000000 to 004027. Each slot backs two files on the CDN:

datasets/items/v1/{slotId}.json   ~240 bytes: id, title, year, image, edges
images/items/v1/{slotId}.webp     at most 1024px wide, quality 76

Both URLs derive from the slot number, so the client fetches the still and the metadata in parallel, with no lookup first. The client knows only the slot count. Random mode draws a slot with crypto.getRandomValues modulo 4,028 and skips repeats within a session. Each slot’s edges name four other slots, baked offline, and the client prefetches them while you play.

The daily round needs no server

Every player on the same local date gets the same five stills in the same order. FNV-1a hashes the date key, prefixed with wyit-daily-v1; the hash seeds a mulberry32 generator, which draws five distinct slots.

2026-11-18FNV-1a hashmulberry321537940887473111(a)
001537001537.json001537.webptitle, year, edgesthe still(b)
Figure 2 The daily round for 2026-11-18. (a) The date, prefixed with a version string, is hashed with FNV-1a; the hash seeds mulberry32, which draws five distinct slots. (b) A slot number is enough to name both files on the CDN, so the still and its metadata are requested together.

The same few lines run twice: in the app and in an inline script in index.html, which works out the first slot and preloads its still before the bundle has parsed. A test pins both copies to the same vectors. Local midnight is the day boundary because that is what Wordle established; with no backend, a global clock would buy nothing. The cost is that scores are client-trust and a device’s time zone can move the puzzle. Progress lives in localStorage under wyit.daily.v1, so a half-played daily resumes where it stopped.

Two requests per round

The app shell is the bulk of a cold first load; after it, a round is two GETs: the slot’s JSON (max-age=300) and its WebP (max-age=31536000, immutable). The WebP size and quality are a cell-network budget, one to two seconds to the first painted still. Fonts are system fonts, so there is no font download.