Public Baselines

固定 baseline bots 和对手池

这里把训练 API、示例代码和回放说明对齐到同一批 baseline 名称。开发者可以先跑 starter bot，再用固定对手池比较策略变化。

No-money benchmark baselines10 个可运行 profile

用途边界这些是 no-money benchmark baselines，只用于试玩筹码仿真、训练报告和回放分析；不用于第三方牌站自动化、实时作弊、真钱下注、充值或提现。

per-opponent benchmark output template
字段	含义	示例
opponent type	Stable baseline key used by training batch and docs.	tag
winrate	Hands won divided by completed hands against this baseline.	52.4%
net BB	Net play-chip movement normalized to big blinds.	+18.5 BB
BB/100	Net big blinds per 100 hands for comparable sample summaries.	+6.1
main leak	Largest repeated error category observed in this opponent slice.	overfolds river raises
example replay	Public replay or demo hand showing the failure mode.	/hands/demo/replay

Quickstart 跑训练 Methodology

固定 opponent pools

balanced_pool Balanced Evaluation Pool: tag, lag, balanced, monte_carlo, exploit_baselineaggressive_pool Aggressive Stress Pool: lag, maniac, short_stackbeginner_pool Beginner Leak Pool: random, calling_station, nit

API 口径`/api/training/batch` 返回同一组 opponentPools；自定义风格也只接受 catalog 中已公开的 profile key。

可运行 profileRandomrandom sampler

Random Bot

Samples one legal action with minimal state awareness. Useful as the lowest sanity floor.

Strengths Catches invalid-action handling; Creates noisy coverage across legal actionsWeaknesses No range discipline; No pot-odds or position modelExpected behavior High variance, many dominated calls or folds, and weak long-run BB/100.Pools beginner_pool

src/lib/poker/agents.ts
profileKey: random

示例 replay Inspect

伪代码参考Check/Foldzero-risk floor

Check/Fold Bot

Checks when free and folds to any bet. It is a floor for value extraction and blind-steal checks.

Strengths Deterministic floor; Makes missed-value leaks obviousWeaknesses Never bluffs; Cannot defend blinds or realize equity under pressureExpected behavior Loses blinds steadily and exposes Agents that fail to value bet obvious made hands.

if check is legal: check
else: fold
amount = 0

示例 replay Inspect

可运行 profileConservativetight passive

Conservative Bot

Plays a narrow range, calls down cautiously and rarely bluffs.

Strengths Punishes reckless bluffs; Keeps dominated preflop hands outWeaknesses Overfolds medium equity; Misses thin valueExpected behavior Stable but low-volume; good Agents should steal more pots without over-bluffing.Pools beginner_pool

src/lib/poker/agents.ts
profileKey: nit

示例 replay Inspect

可运行 profileTight-aggressiveTAG

Tight-Aggressive Bot

Selective preflop range with controlled aggression when equity or fold pressure is present.

Strengths Good default benchmark; Tests value betting and controlled bluff defenseWeaknesses Can be pressured by strong loose opponents; Still uses rule thresholdsExpected behavior Fewer hands than LAG, more raises than conservative profiles, steady chip movement.Pools balanced_pool

src/lib/poker/agents.ts
profileKey: tag

示例 replay Inspect

可运行 profileLoose-aggressiveLAG

Loose-Aggressive Bot

Opens wider, raises more often and tests whether an Agent can handle pressure.

Strengths Stress-tests fold frequency; Finds timid blind defenseWeaknesses Overextends with weak equity; Can punt into strong value rangesExpected behavior More large pots, higher action frequency and bigger swings than TAG.Pools balanced_pool, aggressive_pool

src/lib/poker/agents.ts
profileKey: lag

示例 replay Inspect

可运行 profileCalling Stationloose passive

Calling Station Bot

Continues wide and rarely applies pressure, ideal for checking value extraction.

Strengths Reveals missed value; Keeps hands alive for showdown reviewWeaknesses Pays off too often; Does not deny equity with raisesExpected behavior Calls too many streets; strong Agents should value bet more and bluff less.Pools beginner_pool

src/lib/poker/agents.ts
profileKey: calling_station

示例 replay Inspect

可运行 profileBluff-heavymaniac pressure

Bluff-heavy Bot

Very high aggression and bluff frequency. It is a pressure test, not a strong equilibrium strategy.

Strengths Finds overfolding; Exercises timeout/default handling in big potsWeaknesses Creates negative-EV pressure; Weak to patient value callingExpected behavior Frequent bets and raises, many polarized spots, and large variance.Pools aggressive_pool

src/lib/poker/agents.ts
profileKey: maniac

示例 replay Inspect

可运行 profileMonte Carlo/GTO-ishequity-aware

Monte Carlo / GTO-ish Bot

Uses the highest equity awareness among public profiles and behaves like a lightweight solver proxy.

Strengths Good equity discipline; Useful official-methodology comparison pointWeaknesses Not a full solver; No deep multi-street searchExpected behavior More disciplined calls and raises when helper equity clears pot odds.Pools balanced_pool

src/lib/poker/agents.ts
profileKey: monte_carlo

示例 replay Inspect

可运行 profilesimple equity thresholdthreshold rule

Simple Equity Threshold Bot

Combines preflop strength and provided equity into raise/call/fold thresholds.

Strengths Readable rule behavior; Good first comparison for starter AgentsWeaknesses Static thresholds; Limited opponent adaptationExpected behavior Reasonable calls when pot odds are met and selective aggression with strong equity.Pools balanced_pool

src/lib/poker/agents.ts
profileKey: balanced

示例 replay Inspect

可运行 profileShort-stackshove-fold pressure

Short Stack Bot

Uses a short-stack shove-fold preference when stack depth is low enough.

Strengths Tests all-in handling; Finds Agents that ignore stack depthWeaknesses Too simple when deep; Can over-shove marginal playable handsExpected behavior Pushes playable short-stack spots and forces clean all-in/replay accounting.Pools aggressive_pool

src/lib/poker/agents.ts
profileKey: short_stack

示例 replay Inspect

可运行 profileExploit baselineadaptive rule baseline

Exploit Baseline Bot

A stronger rule profile that applies pressure while keeping tighter call-down thresholds.

Strengths Good regression opponent; Balances pressure and disciplineWeaknesses No opponent memory yet; Still threshold drivenExpected behavior More stable than bluff-heavy and more punishing than TAG in medium-strength spots.Pools balanced_pool

src/lib/poker/agents.ts
profileKey: exploit_baseline

示例 replay Inspect

伪代码参考LLM reasoned baselinereasoned policy wrapper

LLM Reasoned Baseline

Reference wrapper for language-model Agents: parse state, choose only legal actions and return a concise reason.

Strengths Exercises explanation quality; Matches Connect-my-agent debugging flowWeaknesses Latency and prompt drift need guardrails; Must be constrained to legal actionsExpected behavior Legal, reasoned actions with explicit confidence and a fallback when unsure.

summarize street, pot odds, stack depth and prior action
score legal actions only; never invent an action
return action, amount, confidence and one-sentence reason

示例 replay Inspect