Blog · General

How we read a live table: from video to JSON

The path a result takes from the table to the API: video segments, a neural network, the check against the game's rules, and why we discard rather than guess.

Published on October 30, 2026 · 6 min read

The question we get most from people evaluating a live casino data API is not about price or endpoints. It is this: where does this data come from? The answer changes everything, because it defines what happens when the source fails — and anyone buying data to compute something on top of it needs to know that.

At LynxMetrics, the answer is short: we read the table, not another site's results list. This article walks the whole path, from video to JSON, and the three decisions that affect the quality of what you receive the most.

Why not scrape somebody else's list

It is the easiest path, and the problem is not moral, it is technical: you inherit the flaws of whoever is ahead of you. If the source loses a round, you lose it too, and you do not even find out. In a round-by-round comparison we measured that on a well-known site: about 20% of the rounds were not in its own list, and the gaps were visible in the intervals between consecutive records.

Reading the table costs far more work and has one advantage that pays for it: it can be verified. Cards and dice are visible objects. Anyone can rewind the video, look at the frame and say whether the reading is right or wrong. No third party's list allows that.

The path, in five stages

1. The video arrives in pieces

The table's broadcast is delivered in short segments, the way any video player consumes it. We download those segments in parallel and hand them to the reader in order — that combination was one of the biggest contributors to bringing our delay down, because waiting for one segment to finish before asking for the next adds network latency at every step.

2. From segments to frames

Each segment becomes a sequence of frames. We read at a few frames per second, and that is a choice, not a limitation: we measured that raising the rate costs 50% to 100% more processor to gain 150 to 200 milliseconds. In a game whose round lasts over 30 seconds, that change does not pay the bill.

3. The neural network reads what is on the table

In each frame, we crop the regions where the game puts the information and a neural network recognises the content: the rank and suit of every card, in baccarat, or each die's face, in Bac Bo. This is not OCR of text on screen: it is recognition of the physical object, in the position where it lands on the felt.

The game helps more than you would think. The tables have a clear mechanical moment of resolution — the camera closes in on the cards or the dice — and that is when the information is visible, large and stable. It is also the moment when bets have already closed, which lets us announce the round's phase before the result exists, as we explain in real-time results.

4. The check against the game's rules

Here is the part that protects the data most, and it is free: the game has fixed rules, so a misread almost always produces an impossible situation.

In baccarat, the third-card rules are deterministic. A hand with four cards on one side, or with three cards in a situation where the table could not have drawn, does not exist — and a hand that is still going to receive a card is recognisable too. That gives two guarantees: the reading passes a consistency test, and you can tell the hand ended, rather than merely that it stopped changing.

When the check does not add up, the round is discarded. That is the most important decision in the whole system, and it has a price: better to miss a round than to publish a guess. Missing data shows up as a gap and can be recovered; wrong data enters your statistics and warns nobody.

5. The round is published and stored

With the round confirmed, it is published immediately — we do not wait for the camera to pull back to the wide shot, because the result is already decided before that. That single change took our delay from 10.75 to about 4 seconds in Bac Bo. After that the round is stored, and the path from the collector to the database adds a few tenths of a second.

Today more than 90% of the remaining delay happens before the round exists for us: it is the reader making sure of what is on the table. That is why no client-side trick buys those seconds, and why every round delivered over /stream carries the timestamps for each stage, so you can measure this with your own clock instead of trusting our number.

What we watch once it is live

An audit is a photograph; operation is a film. Three things stay on every day:

  • Each round's consistency, against the game's rules, before publishing.
  • The time since the last round. A real table pauses for a few minutes to shuffle or change dealer; silence well beyond that is a problem, not an interval.
  • The intervals in the history. Since every round has an id and a settlement time, an interval outside the usual pattern is visible in a query — the same one we recommend you run on your own data, in storing real-time data.

The result of all this, in one number

Before opening the Lightning Baccarat API, we audited 89.1 minutes of the table in production: 143 rounds published and 143 correct, 129 of them checked against an external source and 14 frame by frame, by hand, because the external source had gaps of about 100 seconds. The full method is in how we check the data.

How to check for yourself

This is the part that matters most, because transparency without verification is just talk:

  1. Open the table at your operator and the API page next to it, and compare the next few rounds.
  2. Download a period with /results and run the interval query: they should sit close to the table's pace, with no unexplained jumps.
  3. Compare the same period with any other source, and look at both sides of the gaps.

The free 2-hour trial gives you a key to do that without paying anything. And if you find a discrepancy, send it to us through contact with the time and what you saw: a divergent round is the most valuable information we can receive, because it is the only one that improves the reader.

Frequently asked questions

How are live casino game results collected?

At LynxMetrics, by reading the table's own video: the broadcast's segments become frames, a neural network recognises the cards or dice in the position where they land, and the reading is checked against the game's fixed rules before the round is published. We do not copy another site's results list.

What happens when the video reading fails?

The round is discarded on purpose. Better to miss a round, which shows up as a gap and can be recovered, than to publish a guess, which enters the statistics of whoever uses the data without warning anyone.

Why is reading the table better than scraping a statistics site?

Because you do not inherit the flaws of whoever is ahead of you, and because it can be verified. In a round-by-round comparison, a well-known site was leaving about 20% of the rounds out of its own list; a card or a die in the video, by contrast, anyone can check by rewinding the frame.

How long does it take between a round happening and the data being available?

About 4 seconds in Bac Bo and 5 in Lightning Baccarat, measured in production. More than 90% of that is the reader confirming what is on the table, and every round delivered over /stream carries the timestamps for each stage so you can measure it yourself.

Read also