Normalisation across tables
A classifier trained on more than a hundred casino layouts maps whatever your table calls its fields onto one internal format, so a session at a live table and a session on a different provider produce comparable numbers.
The project
Built to measure craps, not to beat it.
Craps Black 2.0 is a measurement instrument dressed as software. Behind it sits Apex, a statistical engine trained on millions of recorded outcomes, and a neural sequence model whose only job is to tell an unusual run apart from an ordinary one. This page explains how it was built, what it actually does, and what it deliberately refuses to do.
- recorded outcomes in the reference corpus
- 5.4M
- rolls in the rolling window per session
- 500
- from roll entered to panel redrawn
- 120 ms
- casino table layouts normalised
- 100+
How the project started
It started as a spreadsheet. One of us kept a column of rolls, a column of stakes and a column of minutes, and after four months the spreadsheet said something the memory never had: the good nights were real, and they were also almost exactly cancelled out by the ordinary ones. Nothing about that was surprising as mathematics. It was surprising as a personal fact, which is a different thing entirely.
The spreadsheet had a limit, though. Counting is easy; knowing whether a count means anything is not. A total that has come up nine times in fifty rolls looks remarkable until you work out how often that happens by chance on fair dice — and the answer is: constantly. Every player who has ever felt a pattern has felt exactly this, and no spreadsheet tells you which side of the line you are on.
So the spreadsheet became a model. We collected recorded craps rolls, fitted the expected distribution properly, and built something that would answer the only question worth asking: is this deviation larger than chance comfortably produces, or is it not? Craps Black 2.0 is the fourth rewrite of that answer.
The engine: statistics first
Apex 2.0 begins with formal hypothesis testing, not with pattern matching. For every total, box number, line bet and range, the engine holds an expected frequency derived from the 36 combinations of two dice — six ways to make a 7, one way to make a 2, and everything that follows from it. Each observed count is then scored against that expectation as a standardised deviation, so a gap is reported in standard deviations rather than in the language of hunches.
This ordering matters more than it sounds. A model that learns straight from the data will happily learn the noise, and craps is almost entirely noise. By fixing the expected distribution analytically first, the learned part of the system has nothing left to invent: it can only describe how far the session has drifted from a baseline it was never allowed to choose.
Under the hood
Where machine learning comes in
The learned layer sits on top, and it is deliberately narrow. It does not forecast. It calibrates — it decides how surprising a reading is, given everything the corpus has seen before, so that the panel can tell you when a run is genuinely unusual and, far more often, when it is not.
Adaptive baseline
A rolling recalibration every 50 rolls keeps the reference window travelling with your session instead of lagging behind it, which is what stopped early versions from reporting a deviation that had already closed twenty minutes earlier.
Sequence model for surprise scoring
A small recurrent network reads the ordered history and outputs a single calibrated number: how unusual this stretch is against 5.4 million recorded outcomes. On fair dice the honest answer is almost always “not very”, and the model says so plainly.
Anomaly and integrity checks
The same machinery flags input that does not look like craps at all — a mistyped roll, a duplicated entry, a stretch that drifts implausibly far from any fair pair of dice — so your ledger stays clean enough to be worth keeping.
None of these four components predicts the next number, and none of them could. That is not modesty; it is the structure of the game. What they do is make a session legible while it is still happening.
The software
What you actually get
Apex 2.0 — Totals panel
The 6 to 10 totals furthest from their expected count in the current sample, hot and cold, each with the size of its gap in standard deviations and a plain-language reading of whether that gap is remarkable.
Compass 2.0 — Range panel
The session split into Under 7 (2–6), Seven and Over 7 (8–12), with the running share of each against the 41.67% that each outside range is expected to hold, and how far the current session has moved from it.
The session ledger
Every session you log — bankroll in, bankroll out, minutes at the table — rolled into the number that matters most and is hardest to feel: your actual cost per hour, across months rather than across one memorable night.
Limits you set before you sit down
A loss cap and a time cap, entered while you are still calm, with a visible bar as you approach them and a warning when you cross. The software cannot stop you. It can refuse to let it happen quietly.
Apex and Compass are the names of the two panels inside the software. There is no third tier, no premium engine and no upgrade behind them.
Why it matters
What this changes for the player community
This category of software has a bad history, and it earned it. The market is full of products that promise prediction, publish invented accuracy rates and recruit through affiliate chains that pay for players rather than for results. Everything those products sell rests on a claim that cannot be true, and the people who buy them tend to be exactly the people who could least afford to.
A shared standard for honest numbers
Deviations reported in standard deviations, expected values derived from the 36 combinations of two dice rather than from marketing, and no accuracy percentage anywhere — because no honest version of that figure exists for this category.
Your own data, in your own hands
The ledger belongs to you and exports to CSV in one click. A player who can show six months of real numbers is a player nobody can sell a miracle to.
Limits treated as a feature, not a disclaimer
Loss and time caps sit in the main panel, not in a settings page nobody opens. Seeing 68% of your cap spent, in the moment, does more than any warning printed in a footer.
A player who knows their real cost per hour makes different decisions from one who remembers their best night. That is the whole benefit we claim, and we think it is a larger one than it first sounds.
The honest limit
Craps rolls are independent. The result of one roll carries no information about the next, so there is nothing in the history for any model — ours included — to extract about what comes next. No amount of training data changes that, because the missing information was never there to begin with.
So Craps Black measures, records and warns. It does not bet, it does not advise a stake, and it never predicts a total. Anyone selling you the other thing is selling you something that does not exist.
How we make money
One licence, one payment, priced in your local currency. No subscription that renews quietly, no upsell, no affiliate scheme paying people to recruit players. The free trial needs no account and no card, because it is meant to be a real evaluation rather than a lead capture.
Responsible play
The software includes loss and time limits that you set before you sit down, and it warns you when you cross them.
If gambling has stopped being a choice for you, the tracking is not the help you need — Gambling Therapy · BeGambleAware
Email: Support@crapsblack.com · Contact via WhatsApp · Contact via Telegram