Jump Back In

Finding one episode in thirteen is a search problem, and the textbook answer is to halve the range. Binary search opens at the midpoint, so the very first question describes a moment from several episodes past where you stopped, roughly half the time, before anything at all is known about you. The efficient algorithm spoils the show on question one.

So the walk goes forward from the start of the season instead, two or three episodes at a stride. Every card it shows is from an episode you have already seen, right up until the single card that crosses the line — and that is the episode you are about to watch anyway. Then a second pass fills in the episodes the stride skipped, one at a time, and that pass cannot spoil anything at all: every episode it asks about is either already watched or the next one up.

It costs more questions than binary search, and the bound it buys is structural rather than probabilistic. The most a walk can ever reach past your stopping point is one episode less than the stride. That is a property of the shape, not a number that came out of a test.

Every card offers I clearly remember this, Not sure, and No. The wording is part of the algorithm. Only the first answer advances the walk, and a false yes is the one input that can push a card past your stopping point, so it is written as a high bar on purpose: anything short of certainty should land in "not sure", where it costs nothing but another card.

The card carries no episode number. Numbers invite arithmetic — it's episode four, I think I got past four — and reasoning of that kind manufactures yeses that memory never supplied. Progress is a row of dots for the same reason: a count would tell you how many are left and invite you to finish rather than answer.

Nothing on any screen grades you. No score, no percentage, no correct answer, no reveal. The voice is a friend walking you back through your own memory, interested in what you remember and unbothered by what you don't.

The first person to use it who genuinely did not know where they stopped chose "not sure" for the season, answered "not sure" to three cards from a premiere, and got an answer two seasons before where they actually were. The walk had treated an all-"not sure" episode as a no, locked in the season before it, and offered no way to argue.

Three things were wrong, and all three were the same mistake: undershooting had been treated as free. It is cheap by one episode and expensive by three seasons. Now an episode you are hazy on is unresolved rather than final; the walk asks once whether to keep going and continues at the stride if you say yes, which is also the moment to ask, because you have just seen what a card looks like. The result screen gained an escape hatch — I think I got further than that — and a season that finishes offers to carry on into the next one.

The result is phrased to match: an episode to start from, a fallback beneath it because people stop mid-episode and rewatching one is cheap, and a plain sentence when the memory went hazy rather than a number pretending it didn't.

Cards are written offline from Wikipedia episode summaries, then audited three ways before anything ships: does the moment actually identify the episode it claims, does it give away an outcome, and does it name anyone you have not met yet. Only survivors are shipped, and the audit assumes you have seen only as far back as the stride, not up to the previous episode — an assumption that was wrong once and leaked the ending of the episode a walk had stepped over.

Writing the cards is the cheap half and a budget model does it well enough, at a few cents a season. Judging them is not. Scored against the same labelled set, the cheap model caught 30% of the genuine spoilers with the shipped prompt and 62% with sixteen worked examples, where the bar was 90%. It could reliably find the sentence that gave the game away; it just did not think that it mattered. At about a dollar a season, the audit is the cheapest part of this product to get right, so it stays on the model that gets it right.

Pure white or near-black, no tint, one typeface, and a single deep red that marks the one thing you can do next — used at most once as a fill per screen. The card holds the largest text anywhere on the site and everything else steps back from it. There is no imagery at all, which started as a licensing consequence and turned out to be the right call: the corpus is Wikipedia-only, so there is no poster art to use, and a poster wall would have made it look like the app you watch on rather than the thing you open right before.

Everything works from the keyboard — 1/2/3 or y/u/n — and on a phone the answers sit in the lower half of the screen, within reach of a thumb, because one-handed on a couch is the only posture this is ever used in. Light and dark both follow the system and both are first class. The footer says on every screen that nothing you answer leaves the browser, which is true, and is the kind of claim that is only worth anything where someone can read it.