The Prediction Ledger: Muhammad vs. Bonfim — The Night Our Boldest Call Backfired

The Prediction Ledger: Muhammad vs. Bonfim (2026-06-07)

UFC APEX, Las Vegas. Graded against the resolved results. This column always leads with the misses.

We went 9 for 12 on graded picks in Vegas — a 75% night. Respectable. But a scoreboard is not an alibi, and three fighters we backed lost. Before anything else, here are the three, named, with the confidence our model actually carried into the cage.

Where We Were Wrong

Bryce Mitchell def. Santiago Luna (Submission). This is the one we flagged in our own preview as the night's standout model/market divergence — the model leaned the underdog Luna at low confidence with a +11.7-point edge over the line, its single biggest disagreement with the market on the card. We said plainly it was "a coin-flip where the model leans the underdog," not a play. The market was right and the model was wrong: Mitchell submitted Luna. When we make a point of a divergence going in, we make a point of owning it coming out. The line beat us here.

Tom Nolan def. Fares Ziam (Decision). We picked Ziam (medium confidence, 69.9%) — but notably the market was higher on Ziam than we were, so this was a favorite the model rated less certain than the public did. The unbeaten-in-the-UFC Australian got the decision. We did at least read the shape right: we projected a decision, and a decision is what happened. Wrong man, right method.

Edgar Chairez def. Bruno Silva (Submission). A flat toss-up in our model (50/50), the kind of fight we never tell you we've solved. It fell to Chairez by submission. No sting we didn't sign up for.

The Method Misses

Five fights we called correctly by winner but missed on method — Alessandro Costa (won by KO, not our projected finish), Brendan Allen, Chelsea Chandler, Joanderson Brito, and Ketlen Souza all delivered a different ending than we modeled. It's the same pattern this column keeps flagging out loud: **our model is a much sharper judge of who than of how.** We keep projecting decisions; fighters keep finishing.

Honest Wins

Now, after the misses — and led by the call we're proudest of.

Gabriel Bonfim def. Belal Muhammad (Decision). We made a former welterweight champion the underdog in his own main event and sided with the surging Bonfim — our biggest edge of the night (+13 points), medium confidence. Bonfim won the decision, method and all. The read that Muhammad's two-fight slide was the division moving on, not a blip, held up on the marquee.
Jeisla Chaves (Decision) and Ketlen Souza — our two high-confidence favorites both delivered.
Plus correct winners on Costa, Baraniewski, Allen, McGhee, Chandler and Brito.

The high-conviction end of the card carried us; the flagged coin-flips and our one bold divergence did not. That's the honest split.

Season-to-Date Ledger

Through 19 graded cards in 2026, here's the honest accounting — in two cohorts we keep separate on purpose, because they answer different questions.

Every pick, toss-ups included: 163-77, 67.9%. This is the floor — the number you'd get blindly backing every call we make, coin-flips and all. We show it because hiding it would be dishonest.
The conviction picks — the LOCK and HIGH calls this product actually stands behind: 57-13, 81.4%. Our LOCK tier alone is 13-2 (86.7%); HIGH is 44-11 (80%). This is what the product means when it says it's good — not 'bet every fight,' but 'here are the ones we're sure about.'

We lead with our misses every card, and we won't pretend the floor and the conviction number are the same thing. (For the record, this is our live 2026 forward record — a different measurement from the leak-free backtest holdout, ~70% all-picks, that lives on our methodology page. Live results and audited backtests aren't the same animal, and we won't blend them.) The record compounds in public — miss by miss, win by win.

---

Methodology: picks and confidences are the frozen pre-fight model state (predictions.json), joined to graded outcomes (results.json). Model claims reference ~70% (0.6986 leak-free) accuracy on the 2025-2026 holdout (n=428). This column is an accountability record, not betting advice.