Build journal · Building an AI casino under audit discipline · Episode one
Twelve Games in One Night
On the night of 5 August 2026 an AI wrote twelve casino games and deployed them to a live subdomain before morning, and the only reason any of it is publishable is a rule that was set before the first line was written: no number describing a game may be typed by a human.
The platform had one game in July. Faro Dorado, a five-reel slot, built slowly and carefully over several sessions, with a published return of 94.6% and a self-audit button that re-measures it. The plan on record said no more games — one built properly beat five built quickly, and the compliance argument needed depth rather than a lobby.
That plan lasted until about eleven at night, when it was overridden. By 03:00 there were twelve games, a lobby, and a deployment on the server with a checksum matching the local file.
What was actually built
Twelve games, each with its own mathematics, its own interface, and its own published return. The returns are the interesting part, so here they are with the method that produced each one.
| Game | Basis of the published return | Return |
|---|---|---|
| Andaluz Nights | Exact — all 3,375 combinations enumerated | 95.9704% |
| Cosmic Cascade | Exact — closed-form binomial | 96.4359% |
| Plinko | Exact — binomial over 13 slots | 96.3281% |
| Scratch | Exact — all 512 cards enumerated | 94.9219% |
| Roulette | Exact — 36/37 | 97.2973% |
| Mines | Exact — fair odds × 0.97, constant at every cash-out | 97% |
| Crash | Exact — 0.97/X reached, pays X, at every X | 97% |
| Hi-Lo | Exact — repriced from the visible card | 97% |
| Faro Dorado | Measured — 400,000 spins at page load | 94.6% |
| Baccarat | Measured — 200,000 rounds, per bet | banker ≈ 98.9% |
| Video Poker | Measured — 60,000 hands, heuristic holds | strategy-dependent |
| Blackjack | Measured — 60,000 hands, simplified strategy | rule-dependent |
Eight of the twelve returns are exact. Not certified, not audited by a third party, not estimated over a long run — exact, in the arithmetic sense, because the outcome spaces are small enough to walk in full and the browser walks them when the page loads. Three reels of fifteen stops is 3,375 grids. Three cells of eight symbols is 512 cards. Twelve rows of Plinko is thirteen slots with binomial weights. A laptop does all of that before the lobby finishes rendering.
The other four are measured, and they say so on the page.
The rule that makes AI-built games publishable
An AI can write twelve games in a night. That is now an unremarkable claim, and by itself it is worth nothing — the interesting question is why anyone should believe a single number that comes out of them.
The rule adopted here is narrow and mechanical. Every published return is computed at page load from the same constants the game itself uses to play. Not from a spreadsheet kept alongside. Not from a figure a developer arrived at once and pasted into the marketing. From the paytable and the reel strips and the multiplier array that the running game reads, in the visitor’s own browser, on every load.
This has an awkward consequence, which is the reason it works: the number cannot be made to agree with the intention. If the paytable does not return what the label claims, the label changes, because the label is downstream of the paytable rather than beside it.
What twelve games in a night does not prove
Speed is not the achievement, and treating it as one would be the fastest way to lose the argument this platform exists to make. Production is close to free now. Anybody with a subscription can generate twelve games this weekend, and a good number of operators are already generating something — games, content, reel maths, marketing copy — with no method for verifying what came out.
Volume was never the constraint on a gambling product. Verification was, and remains. Twelve games built in a night with a hand-typed 96% on each would be twelve unverifiable claims, shipped faster. The same twelve with returns computed from their own constants and an audit button on each is something else, and the difference is a methodology rather than a model.
That is also the honest boundary of what an AI did here. It wrote the code and the mathematics. It did not decide that no published number may be hand-typed, it did not choose to keep an error ledger, and it does not sign off on a single legal claim on this site. Those decisions were made by a human with twenty years in gambling risk, and the AI’s own mistakes are logged under his instruction rather than his supervision.
Everything else the games carry
The mathematics is the part that photographs well. The compliance layer is the part that matters to an operator reading this.
Each game sits behind a geo-gate that names the statute rather than displaying a generic refusal — select New Jersey and the lobby blocks citing A5447/S4282, select Oklahoma and it plays with a banner recording that SB 1589 takes effect on 1 November 2026. Session limits actually block play when they expire rather than logging a warning. Self-exclusion cannot be reversed by asking. Reality checks interrupt. There is no purchase path, no coin pack, and no redemption of any kind, which is the reason this platform is lawful in all fifty states including the nine that have banned the sweepstakes model outright.
And every game routes through a presentation engine with two modes. One shows a return below the stake as the loss it is. The other celebrates it, coins and all, the way most of the market does. Both modes run byte-identical mathematics, which is the point of shipping them side by side.
Twelve games exist because a plan was overridden at eleven at night. Eight of them publish returns that are exact rather than asserted, and that part was not improvisation.