Our numbers, published

Model accuracy current as of July 2026; dataset and labeling-pool figures as of August 2026. This page is updated as the numbers move. Measured results only, never projections.

How do we test?

Verdict Fencing is trained and evaluated on 25,224 quality-labeled saber touches from real competition video, each labeled by experienced fencers and referees with the winner and the action. "Quality" means the label was made at the top two confidence levels or was reviewed by the referee validation panel. Accuracy is always reported on held-out footage the model has never seen, grouped so that no bout appears in both training and evaluation.

Who labels the data, and why are they anonymous?

31 fencers and referees label for us, 7 of them international referees, and they have produced 27,906 completed labels across 23,752 saber touches since February 2026. Every label records the winner, the action, and the labeler's own confidence in the call.

Calls that come back low-confidence are routed to a separate review panel: 5 validators, 4 of them international referees, have re-judged 3,258 of the hardest touches blind to the original call. Counting those reviews, 5,993 touches, about a quarter of the labeled set, have been judged independently by two or more people. The remainder carry a single confident label. Blanket double-coverage would be better and we do not have it; the honest description is targeted review of the touches most likely to be wrong.

Our labelers are anonymous by design. They work blind to each other, to the model's prediction, and to whose touch it is, so that seniority, reputation and deference stay out of the training data. In a sport where the referee's identity moves opinions, blind labeling is the only way to keep that influence out of the data. The trade is that we cannot offer you a named expert to trust, so we publish the pool in aggregate instead, straight from the annotation platform, and update it when it moves.

How well does it track the blades? 94.03% keypoint accuracy

The foundation of every call is knowing where each fencer, and each blade, sits in every frame. Verdict Fencing's pose model tracks 16 keypoints per fencer, including the blade guard, midpoint, and tip, at 94.03% keypoint accuracy on held-out competition footage. Blade tips at full lunge speed are the hardest single thing to track in fencing video; this is where most of our engineering effort has gone.

How accurate is the two-light call?

One-light touches are unambiguous. The light decides, and Verdict Fencing just reads the box. The AI's job is the two-light touch, where right-of-way determines the winner. On those:

Confidence matters as much as accuracy: when Verdict Fencing is not sure, it says so rather than guessing. In the app, low-confidence calls are flagged as such, the same way a good referee acknowledges a close call.

Which actions does it name correctly?

Every call comes with the action behind it, and naming the action is harder than picking the winner. These are action-recall figures, meaning how often the model names the action correctly. They are not winner accuracy. The two are different measurements and we do not mix them.

The numbers above are measured on phone footage, because that is what the product runs on. The system has also been evaluated against broadcast video in partnership with a national federation; that work is not ours to publish, but it is the reason we are confident the approach is not overfitted to one kind of camera.

How good is 77.1%, anyway?

Better than it sounds, because there is often no single right answer to be 100% of. Ask two experienced people to judge the same two-light saber touch, blind, and they name the same winner about three times in four. That is the real ceiling, and almost nobody in this sport has measured it.

So we did. We ran a Referee Overlap study: 10 panelists (3 international referees and 7 experienced annotators) each called the same 100 held-out touches. No existing label, no other panelist's answer, no model output. 1,000 independent calls in total. Blindness is what makes the result mean anything: it separates disagreement about the touch from deference to the person calling it.

That last row is the one that changed how we think about this. Panelists were asked to flag any call a colleague could reasonably dispute, and they flagged 2.5% of them, while disagreeing with each other a quarter of the time. The disagreement is not people hedging on close calls. It is people confidently seeing different bouts.

The disagreement also tracks the rules rather than the footage. Panelists agreed 80.9% of the time on attack in preparation, and only 60.6% of the time on reprise. The actions where the FIE text leaves the most room are the actions people split on. That is an interpretation problem, not a video problem, and it is why we are not chasing 100%.

What we are not claiming

• These are not the same measurement. Our 77.1% is the model against a consensus label; 74.9% is one human against another. You cannot subtract one from the other and call the difference a gap. What the study establishes is that the target is not 100%.
• The referees did not do better than the annotators. Among the 3 international referees, pairwise agreement was 70.3%, slightly lower than the rest of the panel. That is three people and three pairs, comfortably inside the noise, and we report it because leaving it out would be the dishonest choice. It is not evidence that credentials make you worse.
• One hundred touches is a pilot. Phase A is seeded and running now: 800 touches, with the quota escalating automatically when the first three panelists disagree on the winner. We pre-registered that rule before looking. When it finishes, the numbers on this page change.

Where we are taking this next

Three further lines of work, described here so you can hold us to them:

• Phase A, at scale. 800 touches instead of 100, which is enough to break the agreement rate down by action and by rating rather than reporting one number for everything.
• Agreement by referee rating. We have built an open call-the-touch quiz that records the respondent's rating, from spectator or fencer up through a local badge to an FIE licence, so we can ask whether higher-rated referees actually converge on the same answer more often. That is an open question in this sport and we have not seen it measured. The quiz has barely been used so far, so there is no result to report yet; when there is, it goes here.
• A number we are deliberately not publishing yet. Our review panel currently overturns a large share of the calls it looks at, but it almost only ever looks at calls the original labeler already flagged as low-confidence. That is a selected sample of the hardest touches in the set, so the rate says nothing about how often labelers disagree in general, and quoting it would be a straightforwardly misleading statistic. We will publish it when we can pair it with a matched random sample.

Study figures as of August 2026.

What can Verdict Fencing not call yet?

• Parry-riposte. Telling a real parry from a blade graze requires knowing exactly when and how the blades touched, and 2D video often cannot settle it. It is the weakest common call we make, we know what the next two gains are, and both are in active development.
• Rare and messy actions. Simultaneous attacks and unusual phrases are underrepresented in training data and called less reliably. These are flagged as low-confidence.
• Poor footage. Verdict Fencing abstains rather than guesses when it can't see the scoring box or track the fencers reliably: severe backlight, heavy occlusion, or a camera angle that hides one fencer.

What about speed?

We're benchmarking end-to-end analysis time now and will publish measured numbers here. The same rule applies to speed as to accuracy: we only claim what we've measured.

Where do the numbers go from here?

Every number on this page is a snapshot, not a ceiling. The dataset grows weekly through our expert annotation team, and the two biggest known gains are in active development. When the numbers change, this page changes.