Home » Table Types
Cash, sit-and-go, MTT — what changes when the format does
Our take on cash tables versus tournaments: the bot you ship into a cash game can be evaluated on EV per 100 hands; the bot you ship into a tournament has to survive its own variance for six hours before you find out if it knew what it was doing. That gap shapes everything from how we configure the engine to how we set the kill switches. Below is the format-by-format breakdown we use internally when planning a deployment.
The short scorecard
| Format | Eval window | Hardest thing | Where bots fail loudest |
|---|---|---|---|
| Cash (6-max NLH) | 10–30k hands | Opponent modelling depth | Static ranges that don't track regs |
| Heads-up cash | 5–10k hands | Exploit-equilibrium balance | Over-bluffing rivers after losing pots |
| SNG (turbo, 9-max) | 200–500 SNGs | Bubble-factor ICM | Calling off light on the bubble vs a short stack |
| MTT (large field) | 500–2,000 MTTs | Stack-depth shifting + ICM end | Mid-stage tightness that throws away CHIP-EV |
Cash tables — the long-form game
Cash is where bots either earn their keep or get caught. The signal-to-noise ratio is the best of the three formats: ten thousand hands of 6-max gives us a usable read on the bot's win-rate per 100. Our cash configs read every reg at the table, hold a rolling exploit map, and lean into the spots where the math says we should. They also fold more rivers than people expect — a cash bot that can lay down two pair to a triple-barrel from a nit is a cash bot that will still be at the table next month.
The audit failure we see most often in cash is too few opp-model slots. A bot with thirty opponent profiles in memory and a brand-new reg pool ends up serving the same five "average reg" profiles to everyone. Under autocorrelation, that's visible. We size opp-model pools to the room's reg count plus a margin.
Sit-and-gos — the ICM grinder
SNGs reward a different engine. The hand-by-hand EV calculation matters less than the bubble-factor multiplier on stack equity. A bot that plays GTO cash poker into the SNG bubble will fold itself out of money and quietly under-perform a much simpler shove-fold chart that respects ICM. We've watched two crews try to port a cash bot into the SNG schedule unchanged. Both bled across a hundred-game sample before the operator pulled them.
What we run instead in SNGs is a hybrid: a cash-style engine in the early levels, switching to a Nash-shove chart inflected by an ICM table once the average stack drops below twelve big blinds. The switch-over rule itself was the part we got wrong the most in our first year — we set it on stack-size alone, ignoring blind level and players-left, and the bot was either too tight three-handed or too loose at the bubble.
Multi-table tournaments — the marathon
MTTs are the format where it takes the longest to know whether the bot is good. A single MTT result is noise. A hundred MTTs is still noise. Five hundred starts to mean something, and even then the variance is brutal. We've shipped MTT bots that printed for two hundred entries and broke even across the next three hundred without the configuration changing.
The behavioural shifts a bot has to handle across an MTT: deep-stack early-game, mid-stack mid-game, ICM-heavy late-game, and final-table dynamics where every elimination changes everyone's pay-jump equity. A single static config across all four phases is the most common MTT mistake we audit. We use four configs and a phase-detector. The phase-detector itself has a CI suite — when it fires the wrong phase, we lose money in a way that's hard to spot until the next morning.