T-RSI Mobile Fast Reflex Engine (Mobile v1)

Ultra-Lightweight, Sub-100µs Autonomous Android Operating System Agent

[!WARNING]

⚠️ EXPERIMENTAL RESEARCH RELEASE — PRELIMINARY BENCHMARK NOTICE

• Verification Advisory: This is an experimental research model. Do not trust benchmark results until independently verified by a secondary source.
• Evaluation Environment & Origin: This release is from Team B, evaluated on an NVIDIA RTX 5090 (32GB VRAM) testbench with 128GB System RAM.
• Replication Requirement: Requires additional independent benchmarks and third-party adversarial verification across expanded Android devices prior to production deployment.

T-RSI Mobile Fast Reflex is an edge-native autonomous Android agent engineered for low-latency smartphone UI automation, touch-target gating, and multiplication-free integer decision reflexes.

Key Performance Metrics

• Peak RAM Footprint: 8.62 KB (agent core heap) / 15.4 MB (total Python runtime).
• Decision Latency: 58.3 µs (sub-60 microsecond reflex, >1,000x faster than cloud LLMs).
• Touch-Target Precision: 98.3% (eliminates false taps on non-interactive decor views).
• Arithmetic: Pure integer additions and subtractions (multiplication-free).
• Standard Benchmark Evaluation (60 Tasks): 88.3% Overall Accuracy
  • AndroidWorld (Google Research): 95.0% (19/20)
  • AndroidControl (Google & CMU): 95.0% (19/20)
  • Mobile-Eval (Alibaba): 75.0% (15/20)

Architectural Pillars

1. Strict Touch-Target Gating: Enforces that only interactable, clickable, and focusable elements receive affirmative scores, penalizing static non-interactive header views.
2. Spatial Screen Region Gating: Disambiguates bottom navigation tabs ($Y > 0.8 \times H$) from top action bar headers ($Y < 0.25 \times H$).
3. Component Role Prioritization: Native integer weighting for Switch, EditText, Button, SeekBar, and TabView.
4. Keypad & Symbol Mapping: Normalizes mobile calculator and dialer operators (+, −, ×, ÷, AC, CLR, •) into invariant ternary states.

Repository Files & Model Weights

• model.safetensors: Quantized 1.58-bit ternary tensor weights ({-1, 0, +1}) in standard Hugging Face Safetensors format.
• weights_packed_1.58bit.bin: 2-bit packed binary representation (264 KB, 4 ternary weights per byte) for Android NDK, microcontrollers, or WebAssembly (WASM).
• config.json: Validated BitNet 1.58b configuration and Team B evaluation hardware metadata.
• vocab.json & tokenizer.json: Word-level tokenizer mapping for 8,192 mobile UI classes, action verbs, and system intents.
• engine.py: Standalone, multiplication-free execution engine with from_pretrained() support.
• adb_controller.py: Standalone ADB controller for Android emulators and physical devices.
• live_interactive_test.py: Interactive CLI REPL and live ADB phone runner.

Quickstart

from engine import MobileFastReflexEngine

# 1. Load model weights directly from Hugging Face Hub
agent = MobileFastReflexEngine.from_pretrained("psikosen/t-rsi-mobile-fast-reflex")

# 2. Execute sub-100µs decision on Android UI candidates
candidates = [
    {"id": "hdr_timer", "text": "Timer", "class": "android.widget.TextView", "clickable": False, "bounds": (42, 180, 232, 258), "center": (137, 219)},
    {"id": "tab_timer", "desc": "Timer tab", "class": "android.widget.FrameLayout", "clickable": True, "bounds": (432, 2150, 648, 2314), "center": (540, 2232)},
    {"id": "tab_alarm", "desc": "Alarm tab", "class": "android.widget.FrameLayout", "clickable": True, "bounds": (0, 2150, 216, 2314), "center": (108, 2232)}
]

action, state, latency_us = agent.decide(
    goal="Switch to Timer tab",
    candidates=candidates
)

print(f"Selected: {action['id']} in {latency_us:.1f} µs [State: {state.name}]")
# Output: Selected: tab_timer in 21.4 µs [State: AFFIRMED]

Running Live on Android Emulator / Device

# 1. Run interactive terminal decision shell
python3 live_interactive_test.py --interactive

# 2. Run live goal on connected phone via ADB
python3 live_interactive_test.py --goal "Open Display settings and toggle Dark theme"