Live numbers

Every number, and where it came from

Nothing on this site is typed in by hand. Each figure below is read from the annotation database or the model release, written into the build, and dated. This is the whole set.

Read from the annotation database and published 28 August 2026. Model figures from verdict-models v1.1.

23,752 saber touches labeled
77.1% winner accuracy, held-out phone footage
74.9% agreement between two human panelists

The dataset

Source: annotation database, live at publish time

LabelsCompleted, valid calls recorded by the labeling pool
27,906
Touches labeledDistinct saber touches carrying at least one label
23,752
Labels in the quality setTop two confidence levels, or reviewed by the referee panel
27,668
Touches in the quality setThe distinct touches those labels cover
23,529
Touches seen by two or more peopleCounting the validation panel. About a quarter of the set
5,993

The people

Source: annotation database. Labelers are anonymous by design; these are the aggregate credentials

LabelersPeople with at least one completed, valid label
31
International referees labelingA subset of the labelers above
7
Validation panel4 of them international referees
5
Touches re-judged by the panelLow-confidence calls, reviewed blind to the original
3,258

The model

Source: the release manifest and the run it names. Full methodology

Winner accuracyAll two-light calls, held-out phone footage
77.1%
Winner accuracy, high confidenceThe top 25% of calls by model confidence
95%
Pose keypoint accuracyHeld-out competition footage, blade included
94.03%
Labels the model trained onA snapshot of the quality set, not the live count above
25,224

Action recall, by action

How often the model names the action correctly. Not winner accuracy

Attack vs. counterattackStrongest, and the most common two-light situation
83.8%
RepriseSolid
64.1%
Attack in preparationMiddling
53.5%
Parry-riposteThe weakest common call we make, and we publish it
47.6%
Attack noUnder-called, and the least represented of the five in training
29.0%

The referee agreement study

Source: the Referee Overlap pilot. 10 panelists (3 international referees), 100 touches, 1,000 blind calls. What it means

Two panelists agree on the winnerAny two panelists, any touch, judged blind
74.9%
Two panelists agree on the actionNaming the action is harder than picking the winner
56.6%
Two international referees agreeThree people, three pairs. Inside the noise
70.3%
Touches where all agreed on the winnerAll 10 panelists
37 / 100
Touches where all agreed on the actionAll 10 panelists
12 / 100
Touches split three waysLeft, right and no-touch all called on one exchange
9 / 100
Calls flagged as arguableBy the panelist who made them
2.5%

Phase A is running now: 800 touches, with the panel escalating automatically when the first three disagree. These figures change when it lands.

How this page updates

A script reads the annotation database and the model release manifest, writes one payload, and rebuilds the site. Every figure here comes out of that payload, so the page cannot drift from the database the way hand-typed copy does. If a figure fails its sanity check, the build keeps the last known-good value and says so in the log rather than publishing a zero.

These numbers move when we publish, not continuously. The date at the top is the honest one.

Watch them move

Early-access members get the updates as they happen.

Free tier at launch · One email when your spot opens — no spam.