Blog · Lightning Baccarat

Lightning Baccarat accuracy: a 143-round audit

The audit we ran before opening the Lightning Baccarat API: 143 rounds checked against an external source and frame by frame, and what we still watch daily.

Published on October 30, 2026 · updated on October 31, 2026 · 6 min read

A results API is worth exactly as much as the trust you can place in each of its rows. If one round in fifty comes with the wrong card, no statistic built on top of it means anything — and the worst part is that the error goes unnoticed, because the number still looks like a number.

So, before opening the Lightning Baccarat API for subscriptions, we ran an audit, and this article explains how it was done, with the method and the result. This is not marketing copy: it is what we measured, and what we keep measuring.

Where our data comes from

LynxMetrics does not read another site's results list. We read the table: the live video of Evolution's Lightning Baccarat, frame by frame, with a neural network trained to recognise every card — rank and suit — where it lands on the felt. When the hand closes, the round is assembled from what is on the table and stored with the cards in the order they came out.

That difference matters for two reasons:

  • There is no middleman to get it wrong or fall behind. A scraper pointed at a third party's page inherits that page's gaps and delays; we only inherit the video quality.
  • Verification is possible. A card is a visible object. You can rewind the frame, look, and say whether it is right or wrong — which is exactly what an audit needs.

The method

We captured 89.1 minutes of the table in production and let the reader publish normally, with no intervention. Result: 143 rounds published. Every one of them was checked, no sampling:

  1. 129 rounds against an external source. We compared round by round with the public API of a well-known live casino statistics service, matching the winner, each side's score and the cards.
  2. 14 rounds frame by frame, by hand. Those 14 were not in the external source, because the reference feed has gaps of about 100 seconds — that is, it loses rounds itself. So that none would be left without a verdict, we rewound the video and checked each one in the image.

The result: 143 out of 143 correct. No swapped card, no wrong score, no phantom round and no missing round in our list.

What was checkedRoundsHow
Against the external source129Automatic comparison of winner, score and cards
Frame by frame14Manual check in the image, for lack of an external reference
Total143143 correct

The part nobody shows: the reference's gaps

Those 14 rounds are the audit's most useful finding, and it is worth explaining why.

When you compare a collector with a reference source, the reflex is to treat the reference as the truth. But the reference here is another piece of software, with the same problems: it also reads a table, it can also skip, and in our test it did skip. If we had accepted only the rounds that existed on both sides, we would have reported "129 out of 129" and hidden precisely the rounds where our list was more complete than theirs.

We had already found the same pattern in Bac Bo, with bigger numbers: in a round-by-round comparison, one of the best-known statistics sites was leaving out about 20% of the rounds in its own list.

How long after the table the data arrives

Accuracy is half of it; the other half is delay. We measured that in production too, from the round's settledAt to the data being stored and visible in the API: today it is about 5 seconds for Lightning Baccarat.

More than 90% of that time happens before the round exists for us, in the reader confirming what is on the table. The collector and the database add a few tenths. One change that won a good part of that time was publishing the round as soon as the result is on the table, instead of waiting for the camera to pull back to the wide shot — the result is already decided there, and waiting for the framing was pure delay.

Anyone who wants the number with no polling interval uses /stream, which delivers at the moment of storage, as we explain in live results. Every round that arrives there carries receivedAt and sentAt, so you can measure this delay yourself instead of taking our word for it.

What we watch every day

An audit is a photograph; operation is a film. What stays in place afterwards:

  • An internal check on every hand. The score read has to be consistent with the cards read, and the number of cards has to obey baccarat's third-card rule. A hand that does not add up is not published — the game has fixed rules, and that is a free safety net, which the article on the rules details.
  • An API health monitor, with the time since the last round. A real table stops for a few minutes to shuffle or change dealer; silence beyond 10 minutes is a sign of trouble, not an interval.
  • Gaps in the history. Since every round has an id and a settledAt, an interval outside the usual ~35 seconds between consecutive rounds is visible, and it is the first thing we look at.

How to check it yourself

You do not have to believe us, and the most direct way to verify is to put the table and the API side by side:

  1. Open the Lightning Baccarat table at your operator and our live page next to it.
  2. Compare the cards for the next few rounds. You will see the result appear here a few seconds after it closes there.
  3. With a key from the 2-hour trial, download that period's rounds through /results and check the settledAt values: the intervals should sit around 35 seconds, with no unexplained jumps. The script is in Lightning Baccarat history.

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

Where does LynxMetrics get Lightning Baccarat results from?

From the live video of Evolution's own table, read frame by frame by a neural network that recognises the rank and suit of every card. We do not copy another site's results list, so we do not inherit middlemen's gaps or delays.

Has LynxMetrics' Lightning Baccarat data been audited?

Yes. Before opening the API we captured 89.1 minutes of the table in production, with 143 rounds published, and checked them all: 129 against the public API of an external statistics service and 14 frame by frame in the image. The result was 143 out of 143 correct.

Why did 14 rounds have to be checked by hand?

Because they did not exist in the external source: the reference feed has gaps of about 100 seconds and loses rounds. Instead of ignoring them, we rewound the video and checked each one in the image, so that no round was left without a verdict.

How would you notice if the reader started making mistakes?

Every hand goes through an internal check: the score has to be consistent with the cards and the number of cards has to obey baccarat's third-card rule. On top of that we monitor the time since the last round and the intervals between rounds in the history, where a gap is visible.

How can I check the data myself?

Open the table at your operator and the LynxMetrics live page side by side and compare the cards for the next few rounds. With the 2-hour trial key you can also download the rounds through /results and check that the intervals between settledAt values sit around 35 seconds, with no jumps.

For adults 18 and over only. LynxMetrics is not affiliated with Evolution or with the casinos that offer Lightning Baccarat, and it only sells data and stats, never predictions or betting tips, because every round is random and past results do not predict the next one.

Read also