From a game link to who was on the pitch, second by second.

Broadcasts replay plays from a second camera, sometimes minutes later. A model watching the video sees two goals. We match the recording's audio against itself to find replayed stretches, then fold what happens in them back onto the original moment.

One recording replayed four stretches 515 to 521 seconds late. About 8 goals sat in the replays, which explained most of an early overcount.

Score overlays freeze, reset or show the wrong thing. We read the game's start and end from the play itself, use whistles and commentary when the recording has them, and never require them. Stretches where the clock is unknown stay unknown: minutes become a range plus one consistent best guess.

One broadcast's score panel has a clock field that always reads 00:00. On another, the score matched 18 of 18 goals, but its operator updated it a median 4.2 seconds after the ball went through.

Substitutes warm up beside the pitch, and players on the pitch leave the frame. A cheap pitch check removes bench players before naming, and lineups carry players through camera gaps instead of benching them.

On the development game the pitch check removed 128 of 153 bench units and 2 of 80 units in play.

Asking who is on the pitch every 90 seconds put many of the questions in timeouts and goal resets, where the answer is ambiguous. Roll-call times now come from detecting when the scoreboard clock stops.

27 of 80 fixed-time roll calls fell in stoppages. The clock-stop detector found all 17 stoppages in the answer key.

Carrying a name along a track is how most of the game gets filled in, and also how most errors spread. Names are now carried only between two looks that agree, and a look is dropped when the rest of its track disagrees.

Fill caused 74 of 142 wrong seconds before this rule. With it, the development windows had 1 wrong second in 702.

Same kit, same headband, similar build. The comparison works through a fixed checklist of features that can't change during a game: skin, hair, build, sleeves, socks, cleats, braces. Any clear difference means two people. "Same" needs several matches, and two bodies in one frame are always two people.

Where the answer key could check it, a named gallery cell was right 94 to 97% of the time.

Earlier versions joined thousands of short track pieces correctly, but the joins landed among other pieces and never on the player's main history, so they earned almost nothing. Recognising players against a fixed gallery at regular moments replaced piece-by-piece joining.

Of 282 correct joins in one test, only 11 reached the player's main identity.

A tracker can follow a referee into a keeper across a camera pan and call it one body. Tracks are now cut wherever the appearance model sees the look of the body change.

About 28% of whole-game track strips held two different people. Cutting on appearance change lowered the swap rate from 6.5% to 1.9% in one window and from 2.9% to 0 in the other.

A keeper's headband goes to the next keeper, but models treated it as proof of identity. Comparisons now use a fixed list of features that can't change mid-game, and a headband or number counts only as supporting evidence.

One assembled person absorbed two keepers' stints, costing 443 seconds; readers cited a headband in about a third of their identity answers.

Shown two crops, a model reliably spots a difference but rarely commits to a match. So differences split people freely, while matches need many agreeing features or a gallery look.

A visible difference was right 32 of 34 times. Only 1 of 122 same-player pairs produced a supported match.

A version built to never join two people without calibrated proof joined almost nobody. The simple catalogue it was meant to beat scored more than three times higher, and recognition against a gallery replaced it.

Value 158 against 533 on the same windows. Its settling step threw away 89.1% of 1.68 million accepted joins.

Meta's SAM 3.1 tracks at most 48 objects at once. With bench, crowd and officials in frame, tracking silently collapsed. Shorter chunks fixed most of it, but SAM was not adopted.

In 10 of 13 chunks, tracked people fell from about 25 to 2.

If two confident reads of "13" and "18" each create a person, the game gains a phantom. Now people are assembled from how they look, numbers count as one kind of evidence, and the number is used to name the person at the end. A number read alone never creates anyone.

In one full-game run, 20 of 56 people with a confirmed number had no playing time at all.

On a wide broadcast shot a number is a few pixels tall, and models answer confidently anyway. Number reads are limited to close-ups at least 100 pixels tall, digits seen in separate looks combine only when appearance ties the looks to one body, and some kits (low-contrast prints, dice-style tilted digits) are handled by appearance alone.

Below 100 pixels, full numbers were right 1 to 3% of the time. One player read as #23 ten times, on digits 8 to 18 pixels tall, was #20.

Duplicate numbers happen, and so do reprints. Because identity comes from appearance first, two people with one number stay two people, and a corrected number keeps the person and their history.

A half-hidden 34 reads as 3, and short reads are more common than full ones, so they win the vote. A short number now gives way to the one longer number that contains it when enough independent views read the longer one.

Truncations like 34 to 3, 99 to 9 and 66 to 6 put 3,357 seconds on the wrong player. A roster would have fixed only 25 of 5,024 wrong seconds, because 3, 9 and 6 were real shirts too.

We tried having a model rewatch film with its own notes. It never changed its mind. Independent looks that never see each other's answers are now the rule, and an earlier pass's questions are used only as places to look.

Across 150 second viewings with notes, the goal list never changed. But first-pass questions fell within 6 seconds of 71% of its misses, against 33% by chance.

Two readings of the same moments moved the score by 154 points. Requiring two separate readings to agree, and checking each look against its own track, turns that noise into abstention instead of error.

Each task gets a small benchmark across several models, and goes to the cheapest one that is good enough on the errors that matter. Argon read jersey numbers best on the development game. In the blind test Sol did every job, at low effort for reads and high effort for recognition. Claude Opus was tested as a second reader and never named a wrong player, but abstained too often to move the result.

A video model counted goals on the same footage under two prompts. Every count now records its prompt, prompts ask what the model saw, and both styles run to give a range.

A prompt mentioning a whistle gave 30 goals; a neutral one gave 24; the truth was 28. Primed with the whistle, it reported 17 goals without seeing the ball go through a hoop.

Each model states confidence on its own scale, so a single threshold means different things for different readers. Two cheap independent passes, with disagreements sent to a third, catch far more.

Where three independent passes agreed on 19 goal questions, all 19 were right; all 7 wrong answers sat in the 4 where they split. Opus said 0.9 or more on 2 to 10% of number reads, Sol on 58 to 84%.

Turning Sol up to its maximum reasoning effort made it slower and less accurate on number reads, so reads run at low effort.

At maximum effort Sol got 0 test cases right and abstained 6 times, against 2 right at low and high effort, in about twice the time.

Every result is scored against a hand-made, second-by-second answer key for stretches drawn at random before anyone looks at them. Code that runs during a game can't open the key; scoring code is kept separate.

Rules chosen after seeing a game's scores are development evidence about that game, nothing more. The fresh test froze and hashed every piece of the method first, preregistered which number counts, and ran the whole thing twice.

All 17 hash checks passed before scoring opened.

One 120-question pilot produced no answers the key could check: its questions fell outside the scored stretches. Every new test now pushes perfect answers through the scorer before a single model call.

Hand-made keys contain mistakes too. When two or more independent sessions agree against the key, the moment goes to a person as a clip question, and both scores are reported.

On the development game, three players' numbers in the key didn't match their shirts. Correcting them moved film-matched players from 12 to 15 of 26.

The fresh game's key writes #00 as "0", so the frozen scorer counts correct seconds as wrong. The official score keeps that by contract; a corrected rescore is reported beside it.

14 correct seconds per run were scored wrong. With the alias fixed, the whole-game identity score moved from −281 to 13.

A change was announced as a big win because the field was assumed to be a failure rate. It was a pass rate. Every claimed improvement is now checked against the scorer's own definition.

"3.0% to 0.33%" was a 9-times drop in passing, not in failure.

Python's isolated mode ignores the fixed hash seed, so two export runs packed questions in a different order and completed answers no longer matched them. Exports now pin the seed, and two model-free exports must match byte for byte.

313 paid model opinions on the fresh game were orphaned.

Gemini Flash's video mode stops after 23 looks with a generic internal error. Budgeting looks in the prompt and using five-minute clips fixed it.

10 of 10 failed jobs had made exactly 23 looks. The same clip then gave 9 goals instead of nothing.

Background drivers outlived a requested pause and kept calling a model, and spending caps were checked one call at a time. A pause is now confirmed by checking running processes, and one lock counts money already in flight.

After a pause, detached drivers made 196 more calls over 2 hours.