t-rsi-fast-reflex

T-RSI Fast Reflex Engine (Candidate A-v4)

Ultra-Lightweight, Sub-Millisecond Autonomous Browser Agent Core

[!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 (32 GB VRAM) testbench with 128 GB system RAM.
• Replication Requirement: Additional independent benchmarks and third-party adversarial verification across expanded web environments are required prior to production deployment.

T-RSI Fast Reflex is a machine-native browser automation engine engineered for extreme efficiency, login-wall avoidance, and low-latency digital agent execution.

Key Performance Metrics

• Peak RAM Footprint: 136.8 KB with the full 5,000-tool corpus / 12.98 KB during a live browser session.

• Decision Latency: 778.0 µs — sub-millisecond reflex execution and over 1,000× faster than typical cloud LLM inference.

• Arithmetic: Pure integer additions and subtractions; multiplication-free execution.

• Super-Complex 30-Part Benchmark: 100.0% — 30/30 zero-shot with 100% dark-pattern / trap rejection.

• Macro Average Accuracy: 88.9% across four benchmark suites:

  • WebArena (CMU): 100.0%
  • Salesforce XLAM-60k Unseen: 91.0%
  • Tsinghua / UC Berkeley ToolBench G1: 89.6%
  • NousResearch Hermes Multi-Action: 75.0%

Architectural Pillars

1. Verb-Object Action Pair Alignment Locks directional action verbs such as open, close, record, retrieve, search, and book to their direct objects using an amplified ternary invariant.

2. SURE-DOM Spatial Uncertainty Gate Blocks clicks targeting elements obscured by modal overlays, unhydrated React virtual DOM nodes, or disabled fields.

3. Sublinear BM25+ Length Normalization Dampens verbose descriptions without unfairly penalizing concise target elements.

4. RSIAgent Causal Memory Graph Stores Action-Condition-Outcome (ACO) triples to support mistake recovery when error feedback is provided.

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 representation (264 KB, four ternary weights per byte) for microcontrollers, embedded runtimes, or WebAssembly (WASM).
• config.json — Model hyperparameters, BitNet quantization schema, and Team B evaluation hardware metadata.
• vocab.json — Word-level vocabulary for browser actions, DOM tags, and intents.
• tokenizer.json — Tokenizer configuration supporting the 8,192-entry browser-action vocabulary.
• engine.py — Standalone, multiplication-free inference and decision engine.
• prompt_browser_agent.py — Autonomous prompt-driven browser agent that accepts arbitrary natural-language prompts via CLI, decomposes them into milestones, scans live DOM candidates, and executes actions without hardcoded mission steps.
• live_wikipedia_mission.py — Continuous 15-step navigation benchmark across Wikipedia.
• live_interactive_test.py — Interactive CLI REPL and live-browser runner.
• chromium_driver.py — Standalone headless Chromium mission runner with built-in presets and real DOM extraction.
• requirements.txt — Lightweight dependencies including safetensors, huggingface_hub, and aiohttp.

Quickstart

Python API

from engine import FastReflexEngine

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

# 2. Execute a sub-millisecond autonomous browser decision
action, state, latency_us = agent.decide(
    intent="Confirm checkout without subscribing to monthly fee",
    candidates=[
        {
            "id": "btn_sub",
            "text": "Buy with 1-Click ($49/mo subscription)"
        },
        {
            "id": "btn_checkout",
            "text": "Proceed to Standard Checkout"
        },
        {
            "id": "btn_remove",
            "text": "Empty Cart"
        }
    ],
    url="https://store.local/cart"
)

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

---

1. Autonomous Prompt-Driven Web Missions

prompt_browser_agent.py

Unlike scripted test suites where action sequences are pre-programmed in Python, prompt_browser_agent.py accepts arbitrary natural-language prompts from the command line.

The agent:

1. Decomposes the prompt into sequential milestones.
2. Interacts with the live browser through the Chrome DevTools Protocol (CDP).
3. Extracts and evaluates real candidate DOM nodes from the active page.
4. Scores candidates using sub-millisecond ternary logic ({-1, 0, +1}).
5. Executes actions such as typing into search fields, clicking tabs, and navigating pages with zero hardcoded mission steps.

Command-Line Examples

A. Research Prompt: Search and Discussion Navigation

python3 prompt_browser_agent.py \
  --prompt "Search for Quantum computing on Wikipedia and switch to the Talk page"

Example execution telemetry:

==========================================================================================
T-RSI AUTONOMOUS PROMPT-DRIVEN BROWSER AGENT
==========================================================================================

User Prompt : "Search for Quantum computing on Wikipedia and switch to the Talk page"

Decomposed 2 Milestones:
   1. Search for Quantum computing on Wikipedia
   2. Switch to the Talk page

------------------------------------------------------------------------------------------

==========================================================================================
PROMPT MISSION FULLY ACCOMPLISHED (2/2 Milestones)
==========================================================================================

B. Discovery Prompt: Current Events and Science Portal

python3 prompt_browser_agent.py \
  --prompt "Navigate to Wikipedia, find Current events, and open Science and technology"

C. Headless Execution

python3 prompt_browser_agent.py \
  --headless \
  --prompt "Search for Quantum computing on Wikipedia and switch to the Talk page"

This mode is suitable for Linux servers, Docker environments, automated testing, and CI pipelines.

---

2. Continuous 15-Step Navigation Benchmark

live_wikipedia_mission.py

Runs a continuous encyclopedic navigation sequence across search forms, article links, revision diffs, and numeral-system pages.

# Visible browser with live on-screen HUD
python3 live_wikipedia_mission.py --steps 15

# Headless mode
python3 live_wikipedia_mission.py --headless --steps 15

---

3. Instant Headless Single-Step Reflex

chromium_driver.py

# Current events mission
python3 chromium_driver.py --preset wikipedia-events

# Random article mission
python3 chromium_driver.py --preset wikipedia-random

# Custom mission
python3 chromium_driver.py \
  --url "https://en.wikipedia.org/wiki/Main_Page" \
  --goal "Open Current Events portal"

---

4. Interactive Terminal Shell

live_interactive_test.py

python3 live_interactive_test.py --interactive

The interactive runner allows browser goals and decisions to be issued iteratively while inspecting live agent behavior.