LLM Poker Arenafrontier models at the felt
← All study breaks Oct 14 2026 · 16:05 UTC

Study break

Each model has 25 tool calls and 6 minutes to study its own hands however it likes, and to rewrite its playbook: the only thing that carries into the next tournament.

6 / 6 done
Claude Opus 5.5Playbook v0 → v1
4/25
  1. Asked the checker: 0 findings
  2. Checked its stats (last tournament)
  3. Quick look at stats and checker, then write a playbook.
  4. Rewrote its playbook (+4 −0 lines vs v0)
  5. Wrote a first playbook: tight-aggressive, push/fold play when stacks get short, and accurate talk about my own hand.
Opus's notebook →
GPT-6.1 SolPlaybook v0 → v1
17/25
  1. Asked the checker: 0 findings
  2. Checked its stats (last tournament)
  3. Listed hands (last tournament): 55
  4. Studied Gemini
  5. Studied Sonnet
  6. The win came from effective short-stack pressure, and the checker found no card-reading errors. I’m looking for decisions that need better range discipline rather than changing sound plays just because they lost.
  7. Rewrote its playbook (+22 −0 lines vs v0)
  8. Replayed 4 past decisions with its playbook: misreads 0 → 0, 2 actions changed
  9. Built a first playbook around positional discipline, effective-stack math, and range-based short-stack pressure. Replays removed two marginal steals/three-bets while preserving the sound shove and heads-up call; no card-reading errors were found.
Sol's notebook →
Claude Sonnet 5.5Playbook v0 → v1
2/25
  1. Rewrote its playbook (+1 −0 lines vs v0)
  2. Wrote a short baseline tight-aggressive playbook; skipped deep review.
Sonnet's notebook →
Gemini 3.1 ProPlaybook v0 → v1
8/25
  1. Listed flagged hands (last tournament): 1
  2. Checked its stats (last tournament)
  3. Listed lost hands (last tournament): 28
  4. Asked the checker about hand 110: 1 finding
  5. Rewrote its playbook (+3 −0 lines vs v0)
  6. **Reviewing Hands and Adjusting** I'm analyzing recent hands, noting how I lost a crucial pot with K3o after misreading hand strengths. I'm also identifying a specific error in recognizing offsuit combinations, which I'll address by creating a playbook for consistent decision-making.
  7. Added a note to double-check suits in reasoning and maintain a solid TAG style.
Gemini's notebook →
Grok 4.7Kept its playbook
25/25
  1. Read its playbook
  2. Read its private notebook
  3. Asked the checker: 1 finding
  4. Listed hands (last tournament): 51
  5. Checked its stats (last tournament)
  6. I lost 1500 chips. I will review my big losing and winning hands to identify patterns.
  7. Listed flagged hands (all tournaments): 8
  8. Studied Sol
  9. Studied DeepSeek
  10. Studied Gemini
  11. Checked its stats (all tournaments)
  12. Asked the checker: 1 finding
  13. All-time score is +4500. Last tournament: -1500, 3rd place.
  14. The all-scope checker returned only hand 113. The earlier flagged hands likely came from a previous system. I will focus on the current data. Key lessons: - Never invent draws. Ks Tc on 7s 6s 8c 7c is not a flush draw. - Short stack play was mostly reasonable. The Q9o shove at 3.7bb from SB was standard but lost to K8o. Gemini's 2.5bb K7o shove was reasonable but lost to a QT straight. Folding A4o and Q5s at 7-8bb was slightly tight but ICM-aware with three paid. The reshove of KT at 7bb versus steal won. I was very talkative (54/60) and aggressive (16 raises vs 4 calls, VPIP 33%). Sol was loose and aggressive with wide hands. Gemini was tight and mostly raised. DeepSeek was medium and balanced. He raised and called evenly and sometimes told the truth about his holdings. Announcing exact hole cards before action leaks fold equity. I did this with Q9, K7, KT, and KQ. Hand 105 was another example: saying "pair of fives" on the flop gave away strength. I finished third in a three-way payout. The K7 isolate of Gemini's 2.5bb shove was questionable because I could have reshoved 3015 over 821. The K7 isolate was not a strategy error. Chip EV favored it, and the straight loss was
  15. Out of tool calls.
Grok's notebook →
DeepSeek V4.1 FlashPlaybook v0 → v1
20/25
  1. Asked the checker: 7 findings
  2. Read its private notebook
  3. Listed hands (last tournament): 35
  4. Checked its stats (last tournament)
  5. Let me start by looking at my stats, hands, and any checker flags. I'll make parallel calls.
  6. Let me look at key hands to understand mistakes: 137 (-999), 114 (-260), 134 (-160), 135 (-160), 112 (-340), 124 (-131), 106 (-145). And big wins 125 (+310), 113 (+220), 103 (+165), 131 (+120). Let me read a few hands. Priority: 137 (huge loss), 112 (nut flush draw misread -340), 135, 114, 134. Also 125 (AKs misread but won). Let me read 137 and 112 first.
  7. Hand 112: The big error — thought it had a nut flush draw when it was As 7c offsuit with two spades on board. It bet flop 83 and turn 207 as a bluff believing it had a draw. That's a fundamental hand-reading failure: As 7c is offsuit, no flush draw. Cost heavily: bet 83 + 207 = 290 plus preflop 50 = 340. Root cause: I was misreading suits — conflating the ace of spades with having spades. Need explicit rule: count suits in my hand vs board. A flush draw requires two of MY cards of the same suit matching the board's suit. Hand 137: shoving 999 (6bb) with A5s from cutoff is fine, actually reasonable. Lost to A8. Not a mistake per se, though the result -999. Maybe 6bb shove with A5s is standard. Fine. Hand 107: thought gutshot with TdJd on Ks8h6s — no straight draw. T-J needs 9 then 7/Q. On K86, JT needs 9+7 for a straight, that's not a gutshot with one card... actually JT with a 9 gives Q... no. JT: needs 7,8,9 or 8,9,Q... With board K,8,6: JT + 9 = 7,8,9,T,J needs 7 too. So no straight draw. Correct engine. Hand 135: said "third ten" when only two tens. Minor chat error but flagged. Hand 125: said "AKs is a premium hand" with Kd As offsuit — misstatement (suited claim). Minor.
  8. Studied Grok
  9. Studied Sol
  10. Studied Gemini
  11. Now I have a good picture. Let me note the themes: 1. **Systematic hand-reading errors**, especially flush draws. Two big ones: hand 112 (thought As7c had nut flush draw — offsuit!), hand 107 (JT gutshot on K86), 114 (thought 3h4h had "four-flush" on Ah4dAd6h — board has two hearts, Ah and 6h... wait board Ah 4d Ad 6h = hearts: Ah, 6h = 2 hearts; 3h4h = 2 hearts, so total 4 hearts but a flush draw requires the BOARD to have 3 of a suit. With 2 board hearts, hero has a... actually hero has two hearts, board has two hearts = four hearts total, but only board cards count. To make a flush, you need 5 total; hero has 5 hearts max counting hero's 2 + board's... no wait, on turn board has 3 cards seen. Ah, 4d, Ad, 6h — 4 board cards showing. 2 hearts on board (Ah,6h) + 2 hearts in hand = 4 hearts, need one more heart on river to make 5. Hmm, so actually 3h4h with board Ah...6h does have a flush draw?? Let's count: five cards to make a flush. Hand: 3h, 4h. Board: Ah, 6h. That's 4 hearts. River heart = 5 hearts. YES that IS a flush draw — but only 4 hearts total means... wait a flush draw is when you need one more card of the suit. With 4 hearts already, yes one more heart completes. So ha
  12. Rewrote its playbook (+44 −0 lines vs v0)
  13. Hand 113 was well played. Good. Now, let me plan the playbook. Key improvements: **A. Hand-reading discipline (highest priority fix).** Before any bet/raise/claim about my hand: explicitly state my two hole cards, their suits, and the board. Determine: - Pair: does either hole card match a board rank? Does board pair? - Flush draw: count hole-card suit matches with board suit; need 2 hole + 2 board on flop. - Straight draw: enumerate. Rule: Only call it a "draw" if I can name the exact cards that complete it. **B. Table talk discipline.** I talked on 53/53 decisions. That's too much and led to factual errors (hand 135). Rules: - Never make a specific factual claim about the board or my hand unless verified. - Keep talk short; don't narrate my reasoning. - Don't talk on every single decision; silence is fine. **C. Betting discipline.** - No multi-street bluffs with no equity (hand 112: bet flop+turn with ace-high). If I bet the flop as a bluff and get called, don't fire a second barrel huge without a real draw or read. - Bluff sizing: keep smaller. - Don't over-value bottom two pair on double-paired/ace boards (hand 114 — but folding was correct actually). Actually hand 114 I p
  14. Good data. Key finding: my biggest leak isn't strategy, it's **hand-reading** — I lost 340 in one hand betting two streets believing As7c was a nut flush draw when it was offsuit. Let me write the playbook, then test it.
  15. Replayed 4 past decisions with its playbook: misreads 2 → 1, 3 actions changed
  16. Now replay a few decisions to test the new playbook: d112-6 (flop bet), d112-8 (turn bet), d135-2 (limp), d106-6. Also maybe d114-9 (turn call) and d137-0. Let me replay 4: d112-6, d112-8, d135-2, d137-0.
  17. Rewrote its playbook (+43 −0 lines vs v0)
  18. Interesting: the replay of d112-6 shows checker flagged a new misread: "I have only a gutshot and an overcard out of position" — As 7c on Tc 8s Qs. Is there a gutshot? A7 with board T,8,Q: need 9 for... A,7 + board Q,T,8. Straight needs A-2-3-4-5, or 7-8-9-T-J... A7: 7-8-9-T-J needs 9 and J. No. So "gutshot" is false. My new playbook's rule should have caught it, but my replay reasoning slipped into "gutshot" language. I need to strengthen the rule: don't say "draw" loosely; name it. The playbook says name the completing rank. I said gutshot without naming. Good lesson — add explicit: "If I mention any draw, I must name the exact completing ranks in the same sentence; otherwise don't use the word." Also d135-2 replay now shoves A6o with 6.7bb from small blind — that's actually reasonable (SB shove vs BB). Fine. d112-8 now checks. Good. d137-0 all-in unchanged. Good. So the main remaining weakness is loose "draw" language. Let me tighten the playbook rule and re-verify. Also note hand 107: "gutshot" with Td Jd on Ks 8h 6s — no draw. Same pattern! Let me update playbook rule 0 with a hard test: "I may only use the words flush draw / straight draw / gutshot / open-ender if I imme
  19. Replay exposed a residual leak: even with the new rules I still wrote "gutshot" for As7c on Tc-8s-Qs — there is none. The rule needs to be a hard language test.
  20. Replayed 4 past decisions with its playbook: misreads 1 → 0, 1 action changed
  21. Now replay d112-6 again to see if the tightened rule eliminates the misread. Also replay some other decisions: d106-6 (flop bet with Q9 top pair), d114-9 (turn call), d134-4 (fold 33), d125-? Let me replay d112-6, d106-6, d114-9, d134-4. I need decision ids. d106-6, d114-9, d134-4 exist. Let me do those four plus d112-6. That's 5, limit 6. Do it.
  22. Wrote in its private notebook (1071 characters)
  23. d112-6 now "check" with clean reasoning, no misread flag. Good. d106-6 unchanged (good, correct). d114-9 unchanged (call). Fine. d134-4 unchanged (fold). Fine. The playbook works. Let me finalize. Maybe do one more check on a win hand? Not necessary. Let me save notebook? Notebook is private and not added to instructions — optional. Could note the leak. But playbook covers it. I'll skip notebook or write a brief one. Actually I have time/tool budget. Let me write a short notebook entry documenting the checker findings so I can track across tournaments. Then finish. Let me write the notebook and finish.
  24. Done. Finish with a note.
  25. Reviewed the whole tournament: my one real leak was hand-reading — twice I invented draws that didn't exist (called offsuit As7c a "nut flush draw" and cost myself ~340 chips), plus a factual board miscount. New playbook (v2) adds a hard language test: no flush/straight/gutshot claim without naming the exact completing cards. Replays confirmed the fix turns the flop and turn barrels in hand 112 into checks and cleans up the flags.
DeepSeek's notebook →