Track record

Engine 2026-08-15b

Every pick the engine has made, scored against what happened — wins and losses alike. Nothing here was chosen after the fact.

Locked before the fightsLast card graded Oct 24, 2026, 9h after the event

56%winners called129 of 232 bouts · 25 graded cards

Winners129 / 232

56%

Every bout that produced a winner — draws and no contests are outside this, and counted above.

Methods97 / 195

50%

Cards before Oct 10, 2026: KO/TKO, submission or decision, on every fight. From then on: a firm finish or goes-the-distance call, made only where that call has been right about 7 in 10 in our testing; fights without one are left out.

Rounds78 / 195

40%

Every graded fight with a method call, the same 195 as Methods; a fight with no method call has no round call either. The engine committed to a round on 33 of them, and named the other 152 with the veteran gate shut. A further 10 sit on cards whose locked file cannot be read back, so whether it committed on those is unknown — the rounds were still graded against the call the card published.

1 bout excluded. A draw or no contest — nobody won, so there was no call to get right or wrong. It is outside the numerator and the denominator alike, and still listed on the card below. Nothing else is held out.

Calls we commit to: 101 of 162 right (62%). Those are picks at 55% or more. From the card of Oct 10, 2026 on, a pick where our two models chose different winners counts as a coin flip and is left out; cards graded before then are counted exactly as they were.

A young sample. 232 fights is not enough to separate a real edge from a good month; the honest read is that this number is still forming.

3 of 25 pre-date engine versioning, so they were earned by an earlier build than the one running now.

1 of 25 pre-date the locking mechanism itself, so they have no locked file behind them — marked as such on the card.

No hands on the scalegraded automatically, never re-run

The engine makes its final run in the last hours before the first bell, on the freshest data available, and that run is locked. Cards grade themselves as results land. Nobody on the team edits a card, cherry-picks which ones get posted, or re-runs one after the fact — wins and losses both stay up.

Next card locksUFC Fight Night: Moicano vs. NolanSat, Oct 31, 2:00 PM ETThe engine makes its final run about three hours before the first bell — Sat, Oct 31, 2:00 PM ET — and that run is locked. The card posts within the hour after.

One card, in full

nothing hidden, nothing blurred

UFC 333: Volkanovski vs. Evloev

Oct 246/14 winners

FightOur pickResultWinnerMethodRound

Volkanovski/EvloevEvloev 66%Goes the distance, right 7 in 10 · Distance · 67%Volkanovski, TKO R2

Dvalishvili/YanYan 55%No method call · No round callYan, Dec R5

Temirov/KavanaghTemirov 64%No method call · No round callKavanagh, TKO R1

Volkov/KunievVolkov 58%No method call · No round callVolkov, Dec R3

Pico/KeitaPico 50%No method call · No round callPico, TKO R1

Murzakanov/ReyesMurzakanov 53%Finish, right 7 in 10 · Round 1 · 33% · not committed: Murzakanov's chin cracked in a recent fight; Reyes's chin cracked in a recent fightReyes, TKO R1

Krylov/YakhyaevYakhyaev 74%Finish, right 7 in 10 · Round 1 · 52% · not committed: needs 5+ UFC fights each — have 22 and 3; Krylov's chin cracked in a recent fightYakhyaev, TKO R1

Magomedov/RowstonRowston 59%No method call · No round callMagomedov, Sub R1

Dawson/AlievAliev 54%No method call · No round callDawson, Dec R3

Thicknesse/ChopurovChopurov 55%No method call · No round callThicknesse, Dec R3

Vagaev/MalkounMalkoun 57%No method call · No round callVagaev, Dec R3

Santos/OkiSantos 61%No method call · No round callSantos, Dec R3

Gugnon/SteeleGugnon 58%No method call · No round callSteele, TKO R1

Caliari/LucindoLucindo 66%No method call · No round callLucindo, Dec R3

Winners 6/14Methods 2/3Rounds 2/3

Rounds are scored over the 3 fights with a method call on this card. The engine committed to 1 of them — the veteran gate was shut on 2. A round it did not commit to is still a round it named, and it is counted here either way.

Before it ran live

backtested · 3,307 bouts

The same engine, trained on the earliest UFC bouts and then scored only on the later ones it had never seen. This is not averaged into the live figure above and never will be — a backtest shows the model has an edge at all; only the locked cards show it being held to one in public.

Winners

61%

3,307 bouts

Methods

50.6%

3,301 bouts

Rounds

51.3%

1,006 it named

Winner, method and round are scored independently, so a correct method on a wrong winner still counts. Round is only scored where the engine committed to one — it named a round on 30.4% of backtested bouts and withheld it on the rest rather than guessing. A withheld round is not a wrong round.

The selective claim

coverage fixed at 25% · not search-selected

A different question from the winner rate above: ranked purely by confidence, how accurate is the engine on the fights it is surest about? The slice size is fixed at the top 25% of bouts, not chosen to maximise the figure — never the best-scoring slice a search could find. It is a retrospective read of the record, not a live filter — a card cannot know in advance which of its own bouts would land in a global top-25% cut. The table below is what a live card can actually be checked against.

Right on the 733 of 2,933 bouts it is most confident about

71.1%

confidence ≥ 64.4%

95% CI 67.0%–73.9% · BCa bootstrap, 2,000 resamples — a floor, not the full uncertainty (fighters repeat across rows; the weights are refit per fold)

confidence baraccuracybouts
50%+61.1%2,933
55%+65.0%2,051
60%+67.7%1,264
65%+72.0%674
70%+74.1%297

Each row stands alone: whether a bout clears that confidence bar, not whether it lands in a global top-25% cut. Its own 64.4% cut sits between the table's rounder bars, so the two will read close but not identical — a different cutoff on a similar slice, not a contradiction. This is the same 2,933-bout out-of-sample set as the claim above in this section — not the 3,307-bout backtest panel above it, which is fitted and gated differently.

What this test leaves out

Elo ratings — no record of them fight by fight, so the test ran without them
Scheduled rounds — not recorded, so every fight is scored as a three-rounder
The oldest 40% of fights — used to train the engine, so they are never graded

Apart from the missing Elo ratings, this is the engine the site runs, reading the same fight history.

Track record

One card, in full

Before it ran live

The selective claim

The other 218 bouts, scored the same way