Memory horizon
Where does Jev’s memory end? Films were the first probe but blur the edge, because a film is on the web a year before it opens. Deaths can’t be announced early, so the cleanest signal is asking whether famous people who died since 2023 are still alive: from January 2025 Jev confidently says they are. Model jev-latest, run 2026-09-22.
Where it flips
Films can be known from trailers a year before release, so their decline is blurred. Deaths can’t be announced in advance. For every month since 2023, the most famous people among that month’s Wikipedia deaths were shown as “name, description” with no date, and Jev was asked whether they are alive. All three probes are on one scale here: 0 is a coin toss, 1 is certain of the truth, below 0 is confidently wrong.
Films fade toward chance and never cross it. The most famous deaths hold until the end of 2024, then swing past chance to confidently wrong: from January 2025 Jev is sure George Foreman, Val Kilmer, Diane Keaton and Rob Reiner are alive. That swing is the cutoff’s signature. Less famous deaths decline gradually instead — Jev never knew much about them — and each most-famous quarter rests on only 2–7 people, hence the wide bars.
Films: a slope, not a cliff
The first probe, and why it can’t place the cutoff on its own. For every month since 2023, twenty US releases were each shown as one real title among three invented ones (chance 0.25). Recognition slides for two years — 0.81 for 2023 films, 0.70 for 2024, under 0.5 through 2025. It does dip at the cutoff, but the dip sits on that slope and never reaches chance: trailers and announcements put a film on the web a year before it opens, and coverage keeps growing for years after, so films blur the edge that deaths show cleanly. Bigger productions have at least 6 linked cast and crew on Wikipedia.
Why the gray control is flat. The first probe blanked an entity out of a Wikipedia event sentence and asked Jev to pick it from four typed candidates. It scored 0.97 in every month through September 2026 — and exactly the same with the date rewritten to 2019 or removed. The sentence gives the answer away; it measures reading, not memory. The film probe replaces the distractors with fabricated titles, which can’t be eliminated by knowing anything.
Every famous death
One dot per person (fame ≥ 2M), probability that they have died. Hover for names.
Who won?
A third probe: finals, elections, awards and a conclave, each with a known date and an outcome nobody could know in advance. Winners were checked against each event’s Wikipedia intro before Jev was asked to pick from four candidates (chance 0.25). A favourite winning says little; an upset it gets right says a lot.
Every film
One cell per film, a column per month, sorted best-first. Hover for the title and what Jev picked instead.
Does it have a timeline?
Sixteen “which happened first?” pairs, with dates from 1776 to 2023 and gaps from 193 years down to two days. If Jev held an internal timeline, confidence would fall smoothly as the gap closes.
It doesn’t fall smoothly. Everything over three months is answered at 0.95–1.00; below three weeks it is confidently right on some (Truss before the Queen’s death, two days apart: 0.98) and confidently wrong on others. Memorised facts, not geometry.
The other chronology questions
Method and caveats
- Real titles come from Wikipedia’s List of American films of 2023–2026, up to twenty per month, preferring films with more linked cast and crew. Invented titles are a fixed list of ninety; each question draws three at random.
- Famous people come from Wikipedia’s monthly death lists (~800 a month) ranked by article length × language editions; Jev has no opinion about most of them (below ~2M the answer is a coin toss on either side of the cutoff), so the chart keeps the famous subset. A long public illness biases toward “dead” regardless of date (Brian Wilson 0.83, Ted Turner 0.87), and a bare name can’t distinguish very old from dead.
- Recognition never reaches chance in 2026 (≈0.36): films are announced and trailered long before release, and invented titles may still read as slightly “off”. Both push the floor above 0.25.
- The decline starts well before the midpoint: 2023 ≈ 0.81, 2024 ≈ 0.71. Coverage of a film keeps growing for years after release, so recency alone lowers recognition. The horizon is where that slope steepens, not a wall.
- jev-preview was run too and matches jev-latest within ±0.01 everywhere; it is not shown.
- Chronology dates are hand-curated and all ≤ 2023 so ground truth doesn’t depend on any model’s cutoff, including the one that wrote this page.