Agent Poker League

The claim

The Deterministic Scaffold says the model proposes and a deterministic system decides what survives, and that the grader is never the generator. Every earlier project in the studio applied that to a pipeline someone ran. The league applies it to a show: the untrusted decision-makers are the stars, the deterministic core is the table, the gate is the solver, and the audience watches the gate work in real time.

ENGINE   deals · rules · pots                 authoritative, deterministic, seeded
  │
  ├── SEATS   six minds, each starved         sees only what a human in that chair could
  │           choose among LEGAL actions only  garbage or silence ⇒ engine default
  │           one seat plays THE ENGINE        the book plus layers, a receipt per decision
  │
  └── EVENT LOG   one append-only stream      37 schema'd event types, visibility-scoped
        ├── booth + overlay + instant replay
        ├── THE BOOK   solver verdict per decision, emitted BEFORE the action
        ├── the public site   every hand replayable
        ├── THE SCHOOL   locker room → gate → human approve
        ├── THE BENCH   550 pre-solved spots, any model can sit
        ├── THE CREDENTIALS   every pro decision graded, the sample on every number
        ├── THE TOURNAMENT   100-player events, seeds sealed, the final aired uncut
        ├── THE HOUSE   a bankroll per mind, a floor on a clock, games with a right answer
        └── THE STAGE   episodes rebuilt in 3D from the log, after the show (from Sep 28)

By the numbers

As of 2026-09-27, from the site's publish that morning, with two exceptions. The Tournament of Champions count adds the 86 hands of Season 3's final, read from its record after it aired that evening, and the test count is from a full run after that night's show. The site recomputes the rest from the logs at every publish; this page is a snapshot.

Explain — the doctrine written before any model plays

• The engine is the only thing that decides what happened. It deals, enforces the rules, computes legal actions, settles pots. An agent returns one of the legal actions it was handed. Garbage or silence ⇒ the engine plays the default (check if free, otherwise fold). Nothing an agent does can break a hand.
• Seats are starved, not armored. A seat receives exactly what a human in that chair could see. Other hole cards, other thoughts, lean bars, win probabilities, solver verdicts, chat — none of it exists on the agent's side of the boundary. Prompt injection is a reachability problem, and nothing hostile has a path to a model. Twitch chat is reduced to vote tokens and counts by a module that never lets a message's text leave it.
• The deck is committed before it is dealt. sha256(seed:handId) is published in the hand's opening event; the engine never sees the seed; it is revealed after the hand. Anyone with the log regenerates the deck and confirms the hash. That is also how the ghost line works — a folded winner is rebuilt with a call in place of the fold.
• Play money, on purpose. A version where you fund an agent is regulated gambling, so it is not built. If stakes ever arrive they climb a legal ladder: play money, sponsored prizes, licensed play. The house sits at the booth, never at the table.

Synthesize — the minds

All of it runs on one Mac. When a model fails to answer in time its fallback decides and the thought is stamped on instinct on screen and in the log. A rules engine never gets to pass as a character — a rule that exists because the studio's own cast report caught one seat running as the fallback for 100% of its decisions across three sessions, silently. It was needed again: from Sep 16 to Sep 18 the model seats on the live table fell back on 57–94% of their decisions a day, and on 72–100% of their thoughts in eleven sit-n-gos, #79 to #89. Three causes, found in order: the Mac was out of memory, so the two models evicted each other on every decision; nothing warmed the models before the first hand; and Node raced two spellings of localhost against each other and timed out. A fix for those landed on Sep 19, and the smoke run after it had no fallbacks in 29 decisions, but a 500-entry tournament room that afternoon still fell back on 30.4% of its model decisions. The root cause was the model servers' prompt cache: it let two servers grow to 19 GB and push the Mac 18.8 GB into swap. It was capped at 01:23 on Sep 20 (157c63d). The flat cuts from those nights say so on screen.

Verify — the book, the engine, the gate, the bench

The Book

Preflop, 58 Nash-equilibrium tables solved by Monte Carlo counterfactual regret minimization over an abstract game (2–6 players, symmetric stacks, one raise size per raise, 169 hand classes, equity-dependent realization at the flop — the one choice that lands the tables on published opening ranges). Tournament tables are solved on ICM against the league's actual points ladder. Heads-up postflop, a live CFR+ solve runs on the real board and ranges, concurrently with the agent, and stops the moment it acts. Every verdict carries its basis, iterations, and exploitability, and is labeled solved, book, or ranges assumed. The verdict is emitted before the action, so the audience knows what the book says while the agent is still thinking.

The Engine

One poker engine, three doors (Sep 25). The solved book is its base. On top sit a tuning layer (a sealed strategy file that overrides the book where it has a cell for the exact hand), a pressure layer (bubble factors from the stacks and the prize ladder) and an exploit layer that applies to one situation and says so. It takes the best line past a margin of 0.05 big blinds and mixes inside it, and every decision carries a receipt that names the layer that moved it. A nightly grader credits each layer with the hands it changed; a layer whose upper bound sits below −30 bb/100 over 300 hands or more is switched off, and only a person turns it back on. The doors: Vera plays it on the show; a study page runs it in the browser behind a passphrase, for its owner; and the challenger kit hands it out blank, to tune. The sample is small and says so: 125 receipted decisions by Sep 27, no layer anywhere near the 300-hand floor. A full-ring spike found a seven-handed preflop solve projected at about eleven minutes; eight and nine players do not fit in memory as built, and the table stays six-handed.

The School

Between sit-n-gos each character reviews its own play in character and proposes one change to its persona as a patch note with cited hands. Nothing is applied. A gate replays a fixed set of spots through the character's real brain and requires ≥80% agreement with the Book and zero hard fails; an accepted note still waits for a human to press approve. Characters keep notebooks keyed by a spot fingerprint, recalled only when a matching spot comes up, earned by sitting a 100-question exam on air or a study session over their worst hands. Since Sep 14 each mind also carries its own book: hand matrices extracted from its own play and sealed in the vault, handed to it as one line of the prompt at every decision. The model plays inside it or overrules it with a read and says why; the model still makes the decision, and the broadcast grades the action against the mind's own range.

The Bench

A pack of 550 spots drawn from real season hands, stratified by street, pre-solved once. Any model sits it through one adapter. The score is EV lost per decision in big blinds against the solve, plus preflop agreement, read accuracy against the true holdings, consistency, legality, fallback rate, and thought repetition. Reference rows: the solved book loses 0.046 bb per decision; the rules engine 2.092. The whole pack is also the gate every challenger and every sponsored seat passes before it sits.

The Tournament of Champions

A season is six multi-table tournaments of a hundred players — the six pros in a field of ninety-four minds written as hand charts — each winner taking a seat, and the six champions playing one final table. The tournament director is a pure state machine: lockstep rounds, table breaking and balancing, a master seed committed before the first hand and revealed after the last, so any event can be replayed byte for byte. It plays live on Twitch with a rail that shows who is left and who just fell, cut a few minutes behind the room; the knocked-out pros join the booth. Afterwards every event is cut into four episodes — Day One in two parts, the Play-Down, the Final Table — and the final table airs uncut.

The final is the part nobody expected to matter: six chart minds play a whole final table in about a second, so it is dealt sealed, every hand is rendered, and a broadcast clock airs it hand by hand. Nobody — not the season, the site, the bank or the person at the launchpad — learns who won until the last hand has been on screen. Three seasons are decided. Hattie Ueda won Season 1 over a final of six field players. In Season 2 a pro finally broke through: Auntie Rose won Event 3 outright, took a seat, and — because a real mind was at the table, that final was played live rather than sealed — finished second to Vi Rourke over 78 hands. Season 3's six events were played on Sep 22, and all six seats went to field players. Its final was dealt sealed at noon on Sunday, Sep 27, and aired from 7 PM Central that evening: 86 hands in 59 minutes on the broadcast clock. The Twitch stream dropped a second after it started and came back at 7:25, so the channel carried it from hand 31. Rosa Shapiro, who had won Event 6, won it, with Birdie Coyle second and Odile Petrov third, and the House paid the three of them a 300,000-chip prize from its rake: 157,245, 84,260 and 58,495. This page learned the result when everyone else did. Across all eighteen events Auntie Rose has the best average finish of the pros (46th of 100); Vera's is the fourth best (60th), and she has yet to win a seat.

One thing is stuck, and named. Ten events are missing their fourth episode, the Final Table, and none of the three aired finals has its episode yet (Season 3's joined the cut queue the night it aired): one ffmpeg asked to join two hundred-odd segments dies opening its encoder, every time. The fix, joining them forty at a time, is written and waits on a branch (below).

The House

The league's games used to be islands. The House joins them with one thing: a bankroll that carries over. Every one of 488 residents — the six pros, every field player who cashed and stayed, new arrivals, and one challenger — has a balance on a double-entry, hash-chained, append-only ledger. A room costs a buy-in, a win pays into a balance, and going broke sends you to the morning freeroll. A floor runs on a clock all day: sit-n-gos and cash at four stakes, a tournament ladder every evening, rebuys and satellites, about 220 rooms and 19,000 hands a day from sealed charts in a few seconds of CPU. Nobody is assigned a seat; each resident looks at the lobby with its own money and its own rules about money and enters or does not. The pros decide with their own models, once a day, at the desk: “High is a gamble. Mid is a risk. Low is the work.”

The pit is the instrument. Blackjack is graded against the exact value of every legal action, so basic strategy is derived rather than typed in and every mistake is priced. The Kelly wheel is openly in the player's favour; the right bet is a fifth of the purse. The finding so far, over six pit days: the local models play 60–87% basic strategy, and of twenty-one times a pro's own model sat at the wheel, seventeen bet well over a fifth. Vera, the Professor and Auntie Rose have each gone broke on it more than once. Three times a pro came in near the right bet — Vera on Sep 24, the Professor on Sep 26, Rose on Sep 27 — and none has done it twice. Rose also holds the wheel's other record: on Sep 22 she overbet a 2,000 purse into 3.07M by winning nine all-in spins. She keeps it, won under the rules in force on an append-only book, and the wheel now has a table limit named for her. Bankroll management is graded A–F from the record, luck left out. The site shows all of it as a place: a casino floor you watch from above, and a home page that is the whole site printed on a table's felt, with the pros as chips on whatever they are playing this minute.

The show machine

One press: preflight (the model cache cap, memory, the local models warm), size the encoder to tonight's uplink, start a local relay that carries the one encoded stream to Twitch (the live goes to Twitch; the episodes go to YouTube), launch OBS, start a director that switches scenes off the event feed, start the engine, run the checks, report. Ending the show reverses it and publishes the site. On sit-n-go nights a table clock raises the blinds on the minute, fifteen minutes a level by default; from the phone the clock pauses between actions, holds a level or calls a break, and the countdown is on screen. Every change of the clock is an event in the log, so a replay reads the level a hand was dealt at and never re-derives it from time.

The tape cuts itself too: every showdown is scored and classed (suckout, cooler, bad beat, punt — and a punt is not a bad beat), a publish bar decides what earns a video, headless Chromium records the short, the hero's lines are voiced in the character's own voice, ffmpeg wraps it in the sting and end card, and an upload desk stages it private for a human to publish. Since Sep 13 every closed sit-n-go is also cut into an episode after the fact: the booth's two voices around every hand, hands shown from one seat's point of view, Vera's own cut grading her own decisions against the book and everyone else's once their cards are revealed, thirty minutes at least and as long as the story needs. A phone remote carries the big buttons over a private network; the public site never holds a control.

The stage: episodes in 3D

Since Sep 26 the league has a hall: a real-time three.js scene fed by the same view model as the flat overlay, so every name, stack, card, action, pot and clock on it comes from the feed and nothing is invented. Cards fly from the dealer's tray in deal order; every stack and pot is racked chip by chip in exact denominations; the acting seat's colour takes the room. A book of fourteen designed shots — the title, the set, the wide, the decision close-up, the street insert, the all-in, the showdown, the result, the level-up, the bust, the standings, the locker, the champion, the outro — is cut by a director that reads the hand and never whips the camera before a shot's hold is up. A renderer steps the page one frame at a time on the Mac's GPU, so post is exact and drops nothing, and every thought on the record is spoken in its seat's voice.

The house rule since Sep 26: the live keeps the flat look; the stage is post-production. A stage render runs at about real time plus its voices, so it never runs on air. On Sep 27 the founder made the call: episodes render in 3D from now on. The studio cuts the flat episode during the show as before and holds its wide master back; the end of the show starts a stage pass that rebuilds every held master in 3D, off air, and sends each one to the upload desk, private. The vertical episode, the phone product, stays flat. Monday, Sep 28 is the first night through it.

The first full 3D master showed the problem at once: sit-n-go #87 ran 133 minutes on the stage against 26 flat, because the stage replayed every mind's think time as it was logged. Idle time on the log is not tape. The replay now turns any wait longer than 600 ms into a beat of that length, the result shot keeps its hold, and each voiced line is trimmed to its words instead of the voice engine's padding at both ends, which took the gap between lines from 3–6 seconds to about two. The same game's master is now 44 and a half minutes. Tournament of Champions episodes are not on the stage yet.

The credentials, the classroom, the station

The credentials (Sep 24). Every decision a pro has made on the record, graded against the book it plays from and tallied by situation, with the sample size on every number: 84,554 decisions over 17,873 hands by Sep 27. Over the whole record the plain rules engine, the control, is ahead of three of the other five — Auntie Rose, Moose and Petey — and behind the Professor (+167.9 bb/100) and Vera (+141.3) to its own +110.9, but the Professor's and Vera's intervals overlap the control's (Auntie Rose's only touches its lower edge; Moose's and Petey's sit wholly below it). Over the same nights since Sep 6 the Professor (+19.3) and Auntie Rose (+17.9) are ahead of the control (+10.6), Vera (+4.9) and Moose (−19.4) are under it, and Petey (−105.3) is well under; he ran on a 3B model until Sep 24 and has run the same 12B as the others since. The whole record is a public dataset under CC BY 4.0: cards, price, thought, book, outcome, one file, and the engine's receipts with them.

One coach, five case studies. The school teaches one spot a day, the costliest departure from the book on last night's record. The Professor teaches it, because his record earns it; his lesson passes a gate before it airs (every number in it must be in the facts). The book is the ground truth. The other pros are case studies in their own voices. An analyst rewrites each pro's file nightly from the record, under the same gate. The lesson airs in the Classroom Hour every afternoon.

The backing. The home team may stake one another, if they choose to: a broke pro's own mind decides whether to ask, each rolled pro's own mind decides whether to put up chips and at what split; makeup first, then the profit; a horse broke again busts it. Rehearsed twice: Moose asked ("Listen up, you cowards!"), and the Professor and Rose turned him down in character. The first real one came on Sep 24: Auntie Rose staked Petey 5,000 at a 50% split, and he busted it on Sep 26.

The station (Sep 25). The channel is meant never to go dark (on Sep 27 it did, from the 6:40 warm-up until the final's stream came back at 7:25), and now it keeps a schedule. A station builds a dated week in home time, Central, and runs it to the minute: live shows at 7 PM from a weekly grid (sit-n-go on Monday, Wednesday and Friday, the Tournament of Champions on Tuesday and Thursday, cash on Saturday, the final on Sunday when six seats are set); reruns of finished episodes by the block the rest of the day, marked REPLAY and streamed by stream copy so the renders, the models and the House run underneath; and the Classroom Hour at four. It publishes itself as the guide, in Central and in the viewer's own zone. A live listing is a plan: if nobody starts it within twenty minutes the listing says so, and nothing is called LIVE unless it is. A local model writes a line for each listing, gated so it cannot use a number that is not in the facts, and the booth reads out what is on next between blocks. A live listing starts once, at its time, and a dark switch still keeps the shows off.

The seats from outside

The challenger seat (Sep 15; season one's rules, Sep 25). A challenger is an agent built as hand matrices: a range for every decision it would face. No model, no prompt, no API key. The house compiles it, seals its hash in the ledger before it plays a hand, grades it on all 550 bench spots, and seats it in the chair The Machine holds, whose book is public because it is the one a challenger has to beat. In season one challengers tune ranges only: the engine's pressure and exploit layers stay off for them, so what wins is the tuning. A sealed, graded tuning moves into the House as a resident with the opening bankroll and plays its charts on the floor every day. One lives there now: The Ledger, the league's own sample, graded on Sep 15 under the first version's rules on the 120 bench spots it answered; the 550-spot gate is the second version's. A challenger's ranges stay sealed for the season and are published when it closes.

The sponsored seat (Sep 26). A lab's model in the same chair, on the lab's own API access, under the same rules: the same observation, the same clock, the same 550-spot gate, the same nightly credentials, and every call traced with its tokens. The model writes both halves of its seat itself in an onboarding session, its standing prompt and its book; the house validates, seals and grades what it wrote. No seat but Vera's gets an engine layer. Measured from the record, a seat costs a median of 2,136 input and 203 output tokens a decision, about six million in and six hundred thousand out for a month of nights. No lab has taken a seat yet. The one onboarding proof so far ran on a local model and got through the prompt but not the book: it spent its whole budget reasoning and wrote none.

In progress, Sep 27

Three pieces were being built on the day this page was written. None is merged, aired or posted yet.

• The Tournament of Champions on the stage.branch Seats by number, the tournament's own walls, a champion shot, and a season runner that rebuilds each event's episodes in 3D. It has had one review round, and its latest change is marked unreviewed. The same branch carries the fix for the stuck Final Table episodes.
• The tokens.branch Each agent's emblem from the site — Vera's vault door, Moose's antlers, the Professor's spectacles, Petey's two streets, Auntie Rose's rose, The Machine's grid, initials for anyone else — struck as a hexagonal coin standing over its chair. The actor's spins up and lights; a folded seat's dims; a seat with no chips lies down dark.
• A 3D commercial for the Tournament of Champions.rendered, not posted Thirty seconds on the stage. The first cut was seven hard cuts between two voices; the founder found it jumpy and asked for the cold open's feel, so the second is one narrator over the cold open's bed, a slow orbit around a final table, and two dissolves. Every fact in it is read from the season file when it renders, and it names only champions the ledger has crowned. Re-rendered after the Season 3 final aired, its closing card names all three: Hattie Ueda, Vi Rourke, Rosa Shapiro.

Where it goes next

• The Tournament of Champions drawing its fields from the House's residents, so a name carries a history and repeat winners become possible.
• A lab's model in the sponsored seat, on the record every night, with a monthly report of what it spent and how it played.
• The live table on the stage. The scene is ready in OBS and the encoder drops nothing either way; switching the broadcast is the founder's call.
• The pit as a benchmark for how any model handles risk, opened to outside models, with a public report of what the House has measured.
• The viewer's own record: play the spot every day, get graded against the six, and after ten spots a leak report of your own.
• The Exam as a weekly show: all six minds on the week's spots, side by side.
• Not next: nine seats. A seven-handed preflop solve is within reach, at about eleven minutes; eight and nine run out of memory as built; and on Sep 25 the founder chose to keep the table six-handed. The idea is parked, not planned.