Proof Center
The model archive — wins and losses.
This page is built live from the same database our picks are saved to — nothing here is typed in by hand. One engine, two jobs: Fight IQ calls a winner in every fight from fighter data alone, and Value flags the fights where the odds disagree with the model. Being right isn't enough, so we also show whether our percentages hold up — when we say 70%, does it hit about 7 times in 10? — and every miss.
Loading latest data…
ArchiveThis page documents a model that is no longer part of the product.
Until September 2026 Cannon Fight Lab published a win-probability forecast for every fight. We tested whether
that forecast beat the price the market was offering, could not show that it did, and took it off the public
product rather than keep publishing it with a disclaimer. Nothing here was deleted — the engine still runs and
is still graded in the open, wins and losses at the same size, because a failed test is the one result a
research site must not quietly remove. The live product is now
Card Lab and Market Lab.
Two records, clearly labeled. The live record is the prospective feed: calls published as each card approached, recorded with a timestamp, their percentage and their price. How strongly any individual call's timing can be proved varies row by row, and for part of the record a sealed copy exists that nobody can edit — the Proof Center grades every row and shows you which is which. Everything else is the replay: we re-ran the engine through history year by year, only ever letting it pick fights it had never seen, and graded it at our closing-price proxy. The replay's first year is excluded from everything published — the engine wasn't calibrated yet that early, and an uncalibrated number isn't one we'd have shown you. We never show a simulation as if it were live.
Want the raw rows instead of the summary? Open the Proof Center — every call with its timestamp, its timing grade, price and result, filterable, losses included.
The live record — our prospective calls, graded
Not a simulation. Every call below was recorded with a timestamp as its card approached, then graded against what actually happened — misses and all. How strongly each row's timing can be proved varies; the Proof Center grades them one by one and shows which are backed by a sealed copy. Pick a model to see its live record.
Loading the locked pick record…
Every card, every pick — graded
Crunching the graded record…
Accuracy by year
Do our percentages hold up?
PredictedActual
Where the model has been wrong
The most recent graded misses. A model that only shows winners is selling something.
Every main-engine model represented in this archive — simulations
Every main-engine model represented in this archive, oldest to newest. Fight-duration and prop research run under their own protocols and appear nowhere on this page. These records are simulations — each model re-run over history, not calls published as a card approached — so we grade each only on fights after it was finished, the ones it never trained on. Where a model's exact cutoff isn't confirmed, we start it at Jan 2025 to be safe, so no model can score itself on data it already saw. The one genuinely prospective record is the live section at the top of this page.
Read the high numbers with suspicion, not pride. The models that score highest here (v5 and v6) do so because our own audit found they leaned on the closing betting line or graded themselves partly on fights they trained on — the exact cheating we walked away from. The engine's lower, honest number is the one we stand behind.
Read the high numbers with suspicion, not pride. The models that score highest here (v5 and v6) do so because our own audit found they leaned on the closing betting line or graded themselves partly on fights they trained on — the exact cheating we walked away from. The engine's lower, honest number is the one we stand behind.
Crunching the full pick history…
Accuracy by year
Straight-up picks (the side the model gave over 50%), grouped by the year the fight happened. The gray tick marks 50% — a coin flip.
Do our percentages hold up?
When the model says 65%, those picks should win about 65% of the time. Gray bar = what the model predicted on average, green bar = what actually happened.
PredictedActual
Favorites vs underdogs
Split by the opening betting line, recorded at pick time. Favorites winning often is expected — the underdog column shows what happens when the model disagrees with the market.
Profit per $100 bet
What you'd have made or lost betting $100 on every pick above each edge, at the opening line. Edge is how much higher our percentage is than what the odds say.
Where the model has been wrong
The most recent graded misses, priced at the opening line on record. A model that only shows winners is selling something.
Engine vs the other models, same test
Everything measured on the identical window — fights from Jan 1, 2025 onward, after every model's training cutoff. Apples to apples.
Did we get a better price than the market ended up at?
The toughest test of a betting model, and not one this page is allowed to answer.
Prospective market-price validation is collecting. CFL measures this under the separately frozen CLV-001 protocol, which owns every condition — how many scored observations it takes, across how many separate UFC events, and how tight the range must be. No CLV figure is publication-approved yet. This page holds no threshold of its own and computes nothing here. The Proof Center will display it only when CLV-001's own publication gate is satisfied. Until then, judge us on the graded record above.
How to read this honestly
- Accuracy alone is not enough. Betting the closing favorite every time gets you to ≈68% at closing odds — for free. Our engine lands around 61% without ever seeing a betting line, and we publish that number instead of dressing it up. Fight IQ's claim is independence; Value's claim is price — it only bets where the model and the market genuinely disagree.
- Loading pick provenance…
- Model output, not guarantees. A 70% pick loses 3 times out of 10. Nothing here is betting advice.
- Samples matter. Anything under ~100 picks in a bucket is noise. We show the counts so you can discount accordingly.
The record is the pitch.Put the model to work on the next card — every pick, edge, and price in one place.
Model probabilities are statistical estimates, not guarantees. Historical performance does not predict future results. Cannon Fight Lab is an analytics publication, not a sportsbook, and nothing on this page is betting advice. 21+ only.