Home » Operator Notes
Notes from the table — what we've watched bots do well and badly
This is the longer-form companion to the front page. We've collected the observations below over roughly eight years of running and auditing NLH fleets. Where a number appears, it came out of our own logs, not a market-report deck. Where a number is missing, we'd rather skip it than invent one.
What good bots do, from the seat next to them
A well-built NLH bot at the felt looks ordinary on the surface and weird only in aggregate. Three things separate the good ones we've audited from the merely working ones:
- Pre-flop ranges shift per opponent within the same orbit. The bot that opens 17% UTG against a tight reg and 22% UTG against a calling station, at the same table, in the same orbit, is doing the work. The bot that opens 19.5% UTG no matter who's behind it is leaving money on the table and showing a uniform-stat signature to anyone who notices.
- Bet sizing carries information about board texture. Two-thirds-pot on a dry king-high flop and one-third-pot on a wet middling board is fine. The reverse is also fine if the rest of the line backs it up. What we never want to see is the same sizing across every flop texture for a given bet category.
- The bot folds correctly to obvious cooler lines from regs. A bot that calls down a triple-barrel from a 12/8 nit on a paired board is going to be very profitable until the day it costs you the room.
What bad bots do, also from the seat next to them
The bad ones we've watched fall into three rough buckets:
- Script-readers. Identical timing per street, identical sizing buckets, identical responses to identical situations. A reg with a notes app catches these inside fifty hands.
- Over-tuned to one ecology. A bot tuned on Asian-club fish data, dropped into a European club where the player pool is half regs, is a slow bleeder. We've seen one fleet lose 38 buy-ins across an evening this way before the operator killed it.
- Recovery-incompetent. The bot disconnects mid-hand, reconnects, doesn't know what street it's on, and goes all-in on what it thinks is a turn but is actually the river. We see this perhaps once a month across the fleets we touch.
What broke last time
The most expensive thing that broke in our shop in the last twelve months was a fleet that didn't notice a room's bet-sizing UI had changed. The pot-fraction buttons shifted from 1/3, 1/2, 2/3, pot to 1/4, 1/2, 3/4, pot. Our bot kept clicking what it thought was "2/3 pot" but the underlying API now treated that slot as "3/4 pot". Three days, twelve seats, a known loss before we caught it on the EV-per-100 chart. The lesson logged: every client-update build of every room we run on gets a fresh button-position audit before we let the fleet back on.
What we've stopped doing
We've stopped running bots on tables we can't observe at human speed for the first hour. Some rooms have hand-history exports, some don't. On the ones that don't, an operator sits and watches every seat for an hour, in person, before the rest of the fleet boots. It's slow. It's caught five misconfigurations this year that the automated checks missed.