AI fencing referee
Verdict Fencing is an AI referee for saber. It reads right-of-way on two-light touches from a single phone camera, tells you who the point belongs to, and shows you the action behind the call. No hardware, no wiring, no argument.
Free tier at launch · One email when your spot opens — no spam.
Accuracy, honestly
It calls 77.1% of two-light saber touches correctly on held-out phone footage. That is every call, including the ones it is not sure about. Here is the rest of the picture, including the parts that do not flatter us.
The honest headline: every call Verdict Fencing makes on footage it has never seen, unfiltered by confidence.
On the quarter of calls it is most confident about, it is right about 95% of the time. It flags the rest as unsure rather than guessing.
One-light touches are unambiguous, because the light decides. The AI's job is the two-light touch, where right-of-way settles it, and that is where every number here comes from. Read the full accuracy report, per action and with the methodology →
Who it's for
Two audiences, one engine. They want quite different things from it.
Club night, two lights, and nobody neutral to ask. Film the touch on the phone already in your pocket and get the call in seconds, with the reasoning attached so the touch teaches something.
Free tier at launch. Coaches can review a fencer's touches across a season. See planned pricing →
A consistent second opinion on two-light priority, a training aid for developing referees, and per-action agreement data across tens of thousands of labeled touches.
Built to assist officials, not replace them, and to abstain out loud when the footage will not support a call. Talk to us about a pilot →
How it works
It tracks both fencers and both blades from one camera angle, reads the scoring lights out of the video, and judges the phrase the way a referee does. Built for the reality of club fencing: one phone, ordinary lighting, no setup.
Any modern phone, landscape, anywhere with a view of the strip. Record the touch. A few seconds is all Verdict Fencing needs.
It tracks both fencers and both blades, reads the phrase, and applies right-of-way the way a referee does.
A full breakdown follows: skeleton overlay, blade trajectories, and the action call in plain language: attack, parry-riposte, counterattack.
What it calls
Every call names the action behind it, and naming the action is harder than picking the winner. These are action-recall figures, not winner accuracy. Different measurements, and we do not mix them.
| Action | Named correctly | Where it stands |
|---|---|---|
| Attack vs. counterattack | 83.8% | Strongest, and the most common two-light situation |
| Reprise | 64.1% | Solid |
| Attack in preparation | 53.5% | Middling |
| Parry-riposte | 47.6% | The weakest common call we make, and we publish it |
| Attack no | 29.0% | Under-called, and the least represented of the five in training |
Saber first, because it is the fastest weapon and the hardest right-of-way problem. If Verdict Fencing can call saber, the rest follows. Foil is next, épée after it. Per-action methodology and known limits →
Clear calls
Two lights and a shrug used to mean a five-minute argument. Now it means tap, get the call, and fence on. Verdict Fencing determines right-of-way from video alone: no wired scoring box, no special hardware.
Explained decisions
Every call comes with the reasoning: who initiated the attack, where the parry landed, when the counterattack arrived. It's the feedback loop of having a certified referee at every club night, for every fencer on the strip.
Coming after launch
Win rate by action. Attack speed over the season. What works against lefties. Verdict Fencing remembers every touch you film, so you and your coach can see patterns no notebook ever caught.
The labeling pool
31 fencers and referees, 7 of them international referees, working since February 2026. They label anonymously, and that is deliberate.
Verdict Fencing learns from human calls, so who makes those calls is the whole ballgame. As of August 2026, the pool has produced 27,906 completed labels across 23,752 saber touches. Every label records the winner, the action, and the labeler's own confidence in it.
Calls that come back low-confidence go 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 set, have been judged independently by two or more people. The rest carry a single confident label. We would rather say that than imply a blanket double-coverage we do not have.
Because a name on a call changes the call. Labelers 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 data. It costs us the named-expert byline that normally signals credibility, so we publish the pool in aggregate instead: how many, what kind, how much, and how often two of them looked at the same touch. The full methodology is on the accuracy page →
Who built it
Verdict Fencing is built by William Spear, a saber fencer who was ranked No. 1 in the United States as a junior, finished sixth at the 2010 Youth Olympic Games, and won an NCAA team title with Columbia in 2015 before holding a US senior ranking until he retired in 2019. He has spent the years since as a software engineer building machine-learning and real-time data systems.
He is building it to give something back to a sport that gave him an education, a career, and most of the people he knows. The longer version, and what he actually worked on →
Limits
Plenty, and you should hear it here rather than find out on the strip.
Parry-riposte is the weakest common call it makes. Telling a real parry from a blade graze means knowing exactly when and how the blades met, and 2D video often cannot settle it. Rare and messy phrases (simultaneous attacks, unusual actions) are underrepresented in the training data and called less reliably. Poor footage breaks it outright: severe backlight, heavy occlusion, or a camera angle that hides one fencer.
In all of those cases it abstains rather than guesses. A call it cannot stand behind is flagged low-confidence, the same way a good referee acknowledges a close one. Everything it cannot call yet, in detail →
FAQ
Any phone from the last ~5 years. Landscape orientation, a sideline view of the strip, normal gym lighting. No tripods, sensors, or wiring required, though a propped phone beats a shaky hand.
Saber, first and for now. It's the fastest weapon and the hardest right-of-way problem. If Verdict Fencing can call saber, the rest follows. Foil is next, épée after it.
No. Blade-tip visualization, the glowing trails you see on World Fencing League broadcasts, draws where the blade went. Verdict Fencing does the other job: it decides who the point belongs to and names the action that won it, from ordinary phone footage rather than a broadcast rig.
No. It replaces the argument at club night when there's no referee to be had. For the touches a certified referee should call, Verdict Fencing is a second opinion and a training tool.
There will be a free tier at launch, so you can film a handful of touches every month at no cost. Paid plans add unlimited calls and analytics. See planned pricing; early-access members get launch pricing locked in.
Early access opens this season. Free tier included.
Free tier at launch · One email when your spot opens — no spam.