This repository contains the experimental engine for Plura, a research-focused VLM (Visual Language Model) designed to audit and outperform GPT-5.2 in specific UI navigation tasks.

Key Modules for Reviewers

• Core Logic: plura_engine.py - The main orchestration pipeline.
• Visual Physics: indexing/visual_physics/ - Modules for "Spectral Saliency" and click refinement (auditing perception artifacts).
• Architecture: indexing/ocumamba_lite/ - Implementation of Mamba-based vision encoders for high-efficiency inference.
• Benchmarking: scripts/gpt52_benchmark_fixed.py - The evaluation harness used to compare speed/cost against SOTA.

Infrastructure

• See scripts/VASTAI_DEPLOY.md for GPU cluster deployment notes.

Note: This is an active research repo. You will see failed experiment scripts (e.g., _v1, _debug) which are preserved for audit trails.

GUI Testing Project

Research and benchmarking for GUI visual grounding models.

Project Structure

• scripts/ - Benchmark and evaluation scripts
• indexing/ - Active Inference GUI grounding implementation
• docs/ - Research reports and documentation
• security/ - Rate limiting and security utilities

Key Scripts

• gpt52_cot_benchmark.py - GPT-5.2 benchmark with Chain-of-Thought reasoning
• gpt_benchmark.py - GPT-4o benchmark script
• active_gui_grounding.py - OcuMamba-Lite + Active Inference implementation

Setup

pip install openai datasets pillow
export OPENAI_API_KEY="your-key"

Running Benchmarks

python scripts/gpt52_cot_benchmark.py

Research

See docs/research_report.md for full analysis.