Six-handed, the felt printed CO on three seats at once. Position walked the
table with nextIn, which steps over folded seats, while the seat count it walked
against still included them — so every muck slid the anchors round and the
labels landed somewhere new. Folding the small blind relabelled it the cutoff.
The two walks are a pair and they are easy to confuse. nextIn asks who is still
in the betting; a fold takes you out of it. Position needs the other question —
who was dealt in — because where you sit is decided when the button moves and
does not change because somebody threw their hand away. So nextDealt, which
skips only the seats that are not in the hand at all, and a note at both of them
saying which is which.
The bots never read this. They use InPosition, which really does want the last
seat still live, and which is deliberately not this function. So the policy is
untouched and the money never moved — the only thing this ever broke was the
badge on the plate, which is precisely why nothing caught it.
TestPositionsDoNotMoveWhenSeatsFold deals six-handed, asserts the table prints
each of BTN/SB/BB/UTG/MP/CO exactly once, then folds the seats out from under it
one at a time and asserts nobody's label moves.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
A review pass, and it found the one that would have cost somebody real chips.
Side pots were only ever cut in runout() — the path taken when the betting
stops because nobody is left able to bet. But a hand reaches a showdown with an
all-in player in it and the betting having finished perfectly normally: a short
stack shoves, two players who still have chips behind call, and then keep
betting past them street after street to the river. Nothing was cut. One pot,
everybody eligible, and the short stack takes the lot — every chip the deep
players put in after they were already all-in, money that could never have been
lost to them. All-in for 100 against two players who each put in 500, and the
best hand collects 1,100 instead of the 300 it was playing for.
Chip conservation never saw it. The chips balance perfectly; they just land in
the wrong seat. And every browser session went through runout(), because a
player shoving is what ends the betting. It took reading the code.
Also from the review: play() dereferenced a table it had just been handed as
null, the top-up button offered chips the wallet could not cover, and the
trainer's ETA was sixty thousand hands optimistic on the first line it printed.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
They are different numbers and the felt was quoting the wrong one. Every raked
pot has chips lifted off it, whoever wins — that has to stay true or the table
stops balancing. But the bots' chips are not real, so a pot a bot wins costs
you nothing, and the counter under your stack was climbing anyway while you sat
there folding.
Rake is now every chip off the table (so the chips conserve) and Paid is the
part that came out of a pot you won, which is the only part that is money and
the only part worth telling you about. A chop costs you half of it. The audit
log takes Paid too: the house's income is what it made off the player, not what
it lifted off a bot.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
Phase 4. Hold'em, and it's the only table in the casino that is a session
rather than a game: you buy in, play as many hands as you like, and leave with
what's in front of you. So the live row spans hands and chips cross the border
exactly twice. Everything in between is inside the engine.
The bots move inside ApplyMove, as UNO's do, which is what keeps poker off a
socket: shove all-in and the flop, turn, river, showdown and payout all come
back in one response, as a script the felt plays back.
The CFR policy the plan called "the single highest-value asset in either repo"
was never read. Not once, in the whole life of the game: the trainer wrote its
info-set keys under IP/OOP and the runtime looked them up under BTN/SB/BB, so
every lookup missed and fell silently through to a pot-odds heuristic. Nothing
looked broken, because a policy miss is not an error. And it was the wrong
policy anyway — ten big blinds deep, trained on a tree where a call always ends
the street, which is not poker. So the trainer is rewritten to play the real
engine through the real reducer, at every stack depth the table deals, and the
trainer and the table now build the key with the same function so they cannot
drift apart again. A test fails if the bots stop finding themselves in it.
Three money bugs, and the tests earned their keep. Chip conservation across a
hundred sessions caught an uncalled bet that minted chips. A var-init ordering
trap meant every card was identical, every showdown tied and every bot believed
it held exactly 50% equity. And the browser caught the rake being silently
zero — the tier said 5 meaning percent, the casino handed it 0.05 meaning a
fraction, and integer division took the house's cut down to nothing.
Claude-Session: https://claude.ai/code/session_013M5nD7PgUboJXoDcYHzpuJ
