Settling a Bet From Video: How a Camera Becomes an Oracle (2026)

TL;DR

Settling a bet from video is not a new idea. Horse racing has judged finishes from a slit-camera image since the 1930s, football certifies goal-line systems to a published tolerance and requires a signal within one second, and the VAR protocol spells out when recorded evidence may overturn a referee and when it may not. Each of these systems has the same skeleton: a rule fixed in advance, a recorded decisive moment, a deterministic decision where the evidence allows one, a human judge where it does not, and a defined outcome when the evidence cannot decide. Live-camera betting markets inherit that skeleton. This post explains the precedents, draws the decision tree, and shows how Lukk, by Adkuu, implements it: a hashed rule before the first bet, signed deciding frames, a deterministic counter or a vision judge, and a void that refunds every stake.

Key takeaways

• Evidence-based settlement has three published precedents an operator can study: the athletics photo finish (40,000 images a second, 25-microsecond resolution at Paris 2024), goal-line technology (certified installations, a 1.5 cm tolerance since 2014, a signal within one second) and the IFAB VAR protocol (intervention only for a "clear and obvious error").
• All three fix the rule before the event (torso, whole ball, four reviewable categories), record the decisive moment, and separate deterministic decisions from judged ones.
• All three define what happens on doubt: a dead heat splits the pool, VAR leaves the original decision standing, and a failed goal-line system is simply not used that day.
• A live-camera market needs the same five elements: a frozen rule, a closed window, a deterministic or judged path, published evidence, and a void policy. Lukk's design publishes all five.
• The operator's due-diligence question is not "is the AI accurate" but "what is recorded, who decides, how fast, and what happens on doubt".

Why settlement is an engineering discipline

A sportsbook grades a football bet against an official result that someone else produced. A live-camera market has no federation behind it: the product itself must produce the result, defend it, and refund it when it cannot be defended. That makes settlement the whole product, and it is why the right precedents are not other casino games but the systems sport built when human eyes stopped being good enough.

Those systems are instructive because their design choices are public: IFAB publishes the goal-line principles inside Law 1 and the full VAR protocol, FIFA publishes the certification process, Hawk-Eye publishes its installation-test accuracy, and South Africa's National Horseracing Authority published a step-by-step account of a dead-heat decision in September 2025. Read together, they are a specification for any camera that is asked to be an oracle.

Three precedents and what they fix in advance

The photo finish. The finish-line camera used at the Paris 2024 Olympics, Omega's Scan'O'Vision Ultimate, is a slit camera with no shutter: it records a one-pixel-wide strip of the finish line up to 40,000 times a second, a time resolution of 25 microseconds, and the composite image is those strips laid side by side, so its horizontal axis is time, not distance. Three cameras run (main, backup and infield), and photocells on the finish posts act as a virtual line. The rule is fixed in advance: the finish is the torso, from the clavicle down, so a foot or an arm does not count. In the men's 100 m final on 4 August 2024, that image separated Noah Lyles (9.784 s) from Kishane Thompson (9.789 s) by five thousandths of a second, about 6 cm at Lyles's top speed of 43.6 km/h. Officials reviewed the Omega photo while the eight finalists waited; per the IOC as cited by heise, the main timekeeper, the main referee and the team leader analyse the main camera and a separate official reviews the backup and infield views.

Goal-line technology. Law 1 limits the system's scope in one sentence: "GLT applies solely to the goal line and is only used to determine whether a goal has been scored." The indication "must be immediate and automatically confirmed within one second", delivered to the match officials "via the referee's watch, by vibration and visual signal, and/or via the referee's earpiece/headset". Systems must meet the FIFA Quality Programme: laboratory tests, field tests and real-game situations, then a final installation test by a test institute in each stadium, an acceptance form signed by the institute and the competition organiser, and a certificate that is valid for 12 months before a retest. The tolerance is published and has been tightened: the 2012 Testing Manual allowed a maximum error of plus or minus 3 cm; IFAB cut that to 1.5 cm in the 2014 manual, in force from April 2014. Hawk-Eye's own account, published on 20 May 2026, says its system shows accuracy of under 1 mm in FIFA installation tests, uses 14 tracking cameras per stadium, and sends a validated, encrypted signal to the referee's smartwatch in under one second. The failure mode is defined too: the referee tests the system before kick-off, and if it does not work as the Testing Manual requires, "the referee must not use the GLT system and must report this to the appropriate authorities."

VAR. The IFAB protocol is a judged path by design, and it is strict about when judgement may intervene. The VAR may assist the referee only for a "clear and obvious error" or a "serious missed incident", in four main categories: goal or no goal, penalty or no penalty, direct red card, and mistaken identity. The referee must always make a decision first; only the referee can initiate a review; the original decision stands unless the review clearly shows a clear and obvious error; and "the final decision is always taken by the referee". Factual questions (offside, the location of an offence) are usually settled by a VAR-only review, while subjective ones go to an on-field review in the referee review area. The protocol says "accuracy is more important than speed" and sets no time limit. On failure, the match goes on: a VAR malfunction does not invalidate a match, and if no qualified replacement is available the match continues without VAR.

Two documented disputes settled by evidence

The 2022 World Cup produced the clearest recent case of a decision that stood only because of what a dedicated camera recorded. Japan beat Spain 2-1 on a goal that many broadcast angles suggested had crossed the byline. On 2 December 2022 FIFA stated that the goal "was checked by VAR to determine if the ball had gone out of play", that "the video match officials used the goal line camera images to check if the ball was still partially on the line or not", and that "other cameras may offer misleading images but on the evidence available, the whole of the ball was not out of play." Spain's coach said he felt "something fishy was going on" in the time VAR took, and pundits demanded the images. Two lessons sit in that episode: the decisive evidence must come from the camera positioned for the rule, not from the most convincing angle, and publishing it promptly is what ends the argument.

Horse racing shows the opposite case, where the evidence could not decide and the rule said what to do. In race 5 at Hollywoodbets Durbanville on 22 September 2025 the judges could not separate Saudi Cat and What A Fortune on the positive photo-finish image. The mirror image appeared to put Saudi Cat marginally in front, so the judges checked it against the third- and fourth-placed horses: the superimposed lines on the mirror image did not line up with their noses, which showed the mirror was not aligned with the winning post. After consulting the stipendiary stewards they used only the positive image, applied rule 65.5 (the judge "shall base his decision exclusively on the HORSE'S nose"), and declared a dead heat. The authority published the reasoning and requested that the mirror be realigned before the next race meeting. Unreliable evidence was discarded rather than reinterpreted, the rule's own tie outcome applied, and the pool was split, which is racing's version of a refund.

The settlement decision tree for a live-camera market

The same logic, written as the sequence a camera-settled market must follow. Every step has a precedent in the table above.

Three design points matter most. The rule is frozen before the first bet, as a photo-finish judge does not decide after the race which part of the body counts. The void is a full refund, not a re-grade, as a dead heat splits the pool rather than guessing a winner. And the evidence is published with the result, not on request; FIFA's 2022 statement came after the argument had started.

How Lukk implements the tree

Lukk, by Adkuu, is built on the sentence "the camera is the oracle", and its published mechanics map onto the six steps.

• Frozen rule. The AI writes each question from what the vision layer sees ("Will a bus cross the junction before the light turns red?") in one of five shapes: race, first-to, count, happen and state-at-close. The rule is hashed before the first bet, and the rule_hash pins the exact settlement contract on every ticket.
• Closed window. Bets close before the window begins, so a person standing at the camera knows no more than a player at home when acceptance stops. Each ticket passes a one-to-three-second acceptance delay with an in-flight veto; an operator sees refusals as reasons such as in_flight_event, betting_closed and oracle_lag.
• Deterministic or judged path. A deterministic counter or tracker settles counts, crossings and orderings; a vision judge settles questions that need a reading of the scene. Players watch the same detection boxes the oracle judges, drawn over the video.
• Evidence published. When the window closes, the deciding frames are locked, signed with ed25519 and published with the result; an evidence replay sits behind every result. Markets settle within seconds of close, and larger payouts carry a short verification hold.
• Void. If the camera cannot tell, the market is voided and every stake is refunded, never re-graded. Post-settlement corrections are rare and audit-driven.
• Delivery. The operator receives each result as a signed settlement webhook and through a pull endpoint, in one of three states: won, lost or refunded.

The mechanics are illustrated step by step at lukk.gg/how-it-works, summarised on the Lukk product page, and set out for integrators in How are live-camera betting markets settled?. The wider category is covered in Live-camera betting explained.

What an operator should ask any camera-settled product

1. Is the settlement rule written and hashed before acceptance opens, and can a ticket be tied to it afterwards?
2. Which frames decide, and do they come from the camera positioned for the rule rather than the most convincing angle?
3. Which questions are settled deterministically and which by a judge, and does the settlement record say which path was used?
4. Is the evidence published with the result, signed, and replayable without asking the supplier?
5. What exactly happens on doubt: a void with a full refund, a re-grade, or a supplier decision?

Adkuu's reading is that a product which answers these five questions in writing has done the engineering; one that answers "our AI is very accurate" has not yet started.

FAQ

How does a live-camera betting market get settled? The rule for the market is fixed and hashed before any bet is accepted; when the window closes, the deciding frames are recorded, and either a deterministic counter or a vision judge applies the frozen rule to them. In Lukk the frames are signed and published with the result, markets settle within seconds, and an unreadable scene is voided with every stake refunded.

Is camera settlement as reliable as a photo finish or goal-line technology? The precedents are the right benchmark because they publish their rules, tolerances and failure modes: a 1.5 cm goal-line tolerance, a one-second signal, a torso rule read from a 40,000-images-per-second strip. A camera market should be judged on the same questions: what is recorded, who decides, how fast, and what happens on doubt.

What happens when the camera cannot tell who won? A well-designed market does what racing does with a dead heat and VAR does with inconclusive footage: it does not guess. Lukk voids the market and refunds every stake in full; a void is never a re-grade.

Can a player check the evidence after a settlement? In Lukk an evidence replay sits behind every result, with the signed deciding frames and the rule hash, so a dispute is an audit rather than an argument. Corrections after settlement are rare and audit-driven.

Sources

• Video Assistant Referee (VAR) protocol — IFAB
• Law 1 - The Field of Play — IFAB
• Goal-Line Technology Implementation — FIFA
• The Millimetre Margin: Why Goal-Line Technology is the Standard for Global Football — Hawk-Eye Innovations
• IFAB tighten goalline tech laws to improve accuracy — Inside World Football
• 0.005 seconds: How high-tech cameras decide gold at the Olympics — heise online
• See Noah Lyles win men's 100m in wild photo finish by fraction of a second — NBC Chicago
• Some images were 'misleading': FIFA defends VAR decision over Japan goal — BreakingNews.ie
• NHA Give An Official Explanation Of Monday's Dead-heat Decision — Turf Talk
• Horse Racing Explained: What Is a Dead Heat? — America's Best Racing

Last verified: October 10, 2026.