Sarah Technical Specs

Layered Decision Intelligence: computer vision, rule-aware strategy, and composition-sensitive EV — orchestrated in real time with full explainability.

{% assign radii = "220|190|160|130|100|75|55|35" | split: "|" %} {% for i in (0..7) %} {% endfor %}

Computer Vision

  • GPU-accelerated inference; FP16 on compatible GPUs
  • 60–120 FPS ingest with NVIDIA GPUs; optimized CPU-only fallback
  • Real-time OCR and card recognition with adaptive ROI calibration
  • Multi-monitor capture with Stability Mode for jitter smoothing
  • YOLO-based detection with OpenCV pre/post-processing
  • Automatic seat and hand extraction feeding the Live Grid

Strategy Engine

  • Basic Strategy with legality checks via Rules Engine
  • Omega II balanced count; Illustrious 18 / Fab 4 deviations
  • Ace-density biased deviations (e.g., 10/11 vs A) via side-count
  • Composition-aware EV per move with tunable Monte-Carlo (1k–200k; default 10k)
  • Kelly Criterion betting with integrated bankroll controls
  • Full rule awareness (H17/S17, DAS, resplits, 3:2 vs 6:5, peeking)

Orchestration & State Management

  • Central coordinator: incremental card integration, discard-tray sync, state snapshots, next-bet calc
  • Evaluates outcomes and emits transcript rows
  • Live Grid serves as real-time ground truth for all hands
  • Seat-by-seat hand lists (main/splits), consistent with Card Ticker
  • Running/True Count + decks/cards remaining pushed to Dashboard via signals

Decision Intelligence & Reasoning Pipeline

  • Shoe state continuously synchronized with Omega II and Power Analysis counters
  • Multi-stage decision flow: Power EV → Omega II deviations → Basic Strategy
  • Legality gating ensures only valid moves (double, split, surrender) are considered
  • Targeted rule-aware overrides (e.g., H17 soft doubles) applied automatically
  • Power engine returns EVs for all legal moves with full reasoning transparency

Performance

  • Typical decision latency: sub-100 ms to low 100s of ms (Monte-Carlo budget & CPU dependent)
  • GPU assists vision; CPU handles decision/EV
  • Efficient full-refresh recount avoids double-counting; incremental ticker updates for smooth UX
  • 4K support, multi-monitor capture, low idle CPU footprint

Interface

  • PySide6 modern desktop UI; Light/Dark theming
  • Live Grid, Power Board, Card Ticker, Game Log integrated
  • Contextual dialogs: Insurance, Game Mode, “SARAH – Game Configuration”
  • Help toolbar (deep-linked topics: Vision Feed, Live Grid, Counting & EOR)
  • “Remember my Rules” persistence; Multi-hand features in Unlimited mode

Explainability, Logging & Transcripts

  • Decision traces showing Power EVs, best move, and timing for every hand
  • Per-round/hand transcript logging with optional CSV export
  • Dashboard exposes RC/TC, decks remaining; logs grid locks, rule changes, config updates
  • Fallback paths clearly logged: Power → Omega deviations → Basic Strategy

Data & System Requirements

  • 100% on-device processing; no gameplay data sent externally
  • Optional CSV transcripts; local profiles and preferences stored securely
  • Windows 10/11 (64-bit) or macOS 12+ (Intel/Apple Silicon), 8 GB+ RAM recommended
  • NVIDIA GPU recommended for max vision FPS; CPU-only supported
  • 1080p+ display (4K supported); ~500 MB disk; Internet only for updates/license
Orchestration & State Management

Central coordinator integrates new cards incrementally, syncs discard-tray fraction with counters, snapshots state, computes next bet, evaluates outcomes, and emits transcript rows.

  • Live Grid serves as real-time ground truth for all hands on the table
  • Seat-by-seat hand lists (main/splits), consistent with Card Ticker
  • RC/TC + decks/cards remaining continuously updated and signaled to Dashboard
Decision Intelligence & Reasoning Pipeline
  • Shoe state continuously synchronized with Omega II and Power Analysis counters
  • Decision order: Power EV (Monte-Carlo) → Omega II deviations → Basic Strategy
  • Legality gating ensures only valid moves are considered + targeted rule-aware overrides (H17 soft doubles)
  • Power engine returns EVs for all legal moves with full reasoning transparency
Explainability & Logging
  • Reason strings & timing logs from Power; strategy labels ("power", "deviation", "basic")
  • Per-round/hand transcripts with structured fields and optional CSV export
  • Fallback paths clearly recorded to ensure full auditability

System Requirements & Expected Performance

Minimum

Baseline experience

  • OSWindows 10/11 or macOS 12+
  • CPU2+ cores (i5 6th-gen / Ryzen 3 1200 / M1)
  • RAM8 GB
  • GPUIntegrated/older dGPU (CPU-only vision supported)
  • Storage~500 MB free
  • Display1080p (100% scale recommended)
  • NetworkOnly for updates/license
Expected:
  • Vision: 20–40 FPS @1080p (ROI-dependent)
  • Decisions: 5k–10k sims/move with modest latency
  • Multi-monitor: Supported; keep ROIs modest

Optimal

Smooth experience

  • OSWindows 10/11 or macOS 12+
  • CPU6+ cores (i7 10th-gen / Ryzen 5 3600 / M1 Pro)
  • RAM16 GB
  • GPURecent NVIDIA/AMD (GPU helps vision; CPU drives EV)
  • StorageSSD/NVMe recommended
  • Display1080p–1440p, 100% scale preferred
  • NetworkUpdates/license; local profiles & data
Expected:
  • Vision: 45–90 FPS @1080p; smooth Live Grid
  • Decisions: 10k–20k sims/move typically <100–200 ms
  • Multi-monitor: Stable; use Stability Mode for large 4K desks

Optimal with Vision

High-FPS vision + low-latency strategy

  • OSWindows 10/11 or macOS 12+
  • CPU8+ high-perf cores (i7/i9 12th-gen+ / Ryzen 7/9 5000+ / M2 Pro+)
  • RAM16–32 GB
  • GPUNVIDIA RTX-class (e.g., 3060/3070+), FP16 supported
  • StorageNVMe SSD
  • Display1080p–1440p high-refresh or 4K
  • NetworkUpdates/license only
Expected:
  • Vision: 60–120 FPS @1080p with GPU acceleration
  • Decisions: 20k–50k+ sims/move while responsive
  • Multi-monitor: High-FPS capture; Stability rarely needed
Apple Silicon: Vision: 40–90 FPS @1080p on M2/M3 (MPS acceleration; ROI & scale dependent). Decisions: 10k–20k sims/move typically sub-100–200 ms. Increase Monte-Carlo simulations in Sarah Config for more precision. 4K reduces FPS.

Helpful Notes

  • GPU vs CPU: GPU accelerates vision (detection/OCR). Strategy/EV is CPU-bound and scales with the Monte-Carlo budget.
  • Resolution matters: Higher resolutions increase capture load — shrink ROI or enable Stability Mode if needed.
  • Laptops (hybrid graphics): Set Sarah to use the discrete GPU (Windows Graphics Settings → High Performance).
  • Privacy: No data leaves the machine. All recognition, strategy, and logging run locally.

* FPS and decision latency vary with resolution, capture settings, simulation budget, and hardware. Default Monte-Carlo budget is 10k simulations per move (configurable).