Skip to content

Latest commit

 

History

History
76 lines (61 loc) · 2.92 KB

File metadata and controls

76 lines (61 loc) · 2.92 KB

Changelog

All notable changes to NeuralGTO are documented here.

[1.0.0] — 2026-03-01

MVP Launch — Open-Source Neuro-Symbolic GTO Trainer

Core Pipeline

  • Complete 5-step neuro-symbolic architecture: parse → solve → extract → advise
  • Postflop GTO solving via TexasSolver CFR engine (all board textures, all streets)
  • Preflop GTO lookup via pre-solved ranges (GTO Nexus data, 6-max common spots)
  • Graceful fallback: solver unavailable → LLM-only mode (never crashes)
  • Three analysis modes: fast (LLM-only ~10s), default (2% ~1-2min), pro (0.3% ~4-6min)

GTO Explanation Layer

  • Explains why each action is correct (blockers, range balance, board texture, equity realization)
  • Trust badges on every response (solver-verified vs LLM approximation)
  • Table-ready heuristics — one-line "Table Rule" for live play
  • Sanity checker — LLM reviews extreme solver frequencies

Exploitative Play

  • Opponent tendency override — describe villain, get adjusted strategy
  • Live game prep mode — pool-level exploitative adjustments
  • Always shows GTO baseline first, then deviation

Multi-way Support

  • Pairwise HU decomposition for 3+ player spots
  • LLM synthesis with multi-way theory (MDF compression, sandwich effect)
  • Evaluated on PokerBench: 74.5% accuracy on 424 multi-way preflop scenarios

Study Tools

  • Quiz mode with scoring engine (48 tests)
  • Spot frequency data and study prioritization
  • Hand history import (PokerStars, GGPoker, Winamax)
  • Session history and logging
  • Beginner/Advanced output modes

CLI Features

  • Interactive REPL with conversational gap-filling
  • Range visualization (13×13 ASCII grid via Rich)
  • Rich-formatted output (colors, tables)

Web Interface

  • Streamlit web UI with progress tracking
  • Real-time solver status updates
  • Account system with free-tier gating

Security

  • Input sanitization (14 injection pattern regexes)
  • Rate limiting (per-session, global, per-IP)
  • Brute-force login protection (5 attempts / 15-min lockout)
  • Email validation (format + disposable blocklist + DNS MX)
  • Daily budget tracking (API cost control)
  • Solver binary path validation (prevents arbitrary execution)

Testing

  • 425+ offline unit tests (no API key or solver binary needed)
  • Covers: pipeline, preflop lookup, hand history, multiway, quiz, security, evaluation

Evaluation

  • PokerBench framework integration
  • GTO Lookup: 88.5% accuracy on 244 HU preflop scenarios
  • Gemini Direct: 86.5% baseline
  • Pairwise LLM: 74.5% on 424 multi-way scenarios

Known Limitations

  • TexasSolver is heads-up postflop only (multi-way uses pairwise decomposition)
  • Windows solver binary only (Linux binary available from TexasSolver releases)
  • ~1-3 min per postflop spot (not real-time)
  • SHA-256 password hashing (bcrypt upgrade planned for Phase 2)

Dependencies

  • Python 3.10+
  • Google Gemini API key (free tier works)
  • TexasSolver v0.2.0 binary (optional — LLM-only mode works without it)