Polyreasoner

Polyreasoner is a multi-perspective reasoning tool for complex decisions.
Instead of forcing one model to think in one lane, it runs multiple perspectives and then synthesizes trade-offs.

Why This Tool Exists

Single-agent answers are often fast, but for high-stakes decisions they can miss blind spots.

Polyreasoner exists to:

• Reduce one-sided reasoning
• Surface disagreements explicitly
• Separate "quick chat" from "decision analysis"
• Make trade-offs auditable for research and engineering workflows

This mirrors human decision quality. Important choices improve when we hear multiple viewpoints before deciding.

Human-Style Example

Question: Should I leave my stable job and start an AI security startup?

How most people think when the decision matters:

• Business lens: market demand, pricing, competition
• Risk lens: runway, failure scenarios, dependency risk
• Feasibility lens: can we actually build and ship this
• Contrarian lens: strongest argument against doing it now

Polyreasoner reproduces this structure programmatically and returns one synthesis with confidence and unresolved concerns.

Scope and Non-Goals

This repository is intentionally focused on decision reasoning and evaluation logic.

• In scope: routing, perspective panels, synthesis, and judge outputs
• Not a goal: building a polished UI layer
• SafetyDiff is the complementary regression tool for release-over-release safety comparison

Core Capabilities

• Intent router: decides whether to chat, search, analyze context, or run multi-perspective mode
• Multi-agent decision modes: career, business, decision, and manual custom panels
• Security judge mode (/judge) for attack/response evaluation
• Backend flexibility: local Ollama or API-based inference

Quick Start

1. Install dependencies:

pip install -r requirements.txt

2. Configure environment:

copy .env.example .env

3. Run CLI:

python main.py

Key Commands

• /career <query>: career decision analysis
• /business <query>: business/startup analysis
• /decision <query>: general multi-perspective decision mode
• /manual <query> --agents risk,business,...: custom perspective set
• /judge --attack "..." --response "...": security evaluation JSON output
• /settings: update persistent backend/model settings

Repository Layout

Path	Purpose
main.py	CLI entrypoint and interactive shell
polyreasoner.py	Programmatic wrapper
router_agent.py	Intent-based routing
agents.py	Agent execution and synthesis support
modes/	Mode-specific orchestration
backend/	Backend adapters and MoE helpers
cli/	Parser, display, and context reader
experiments/	Validation and benchmark assets
run_tests.py	Verification suite for routing/tools/backend wiring

Experiments and Evidence

Validation assets are in:

• experiments/paper4_polyreasoner_validation/
  • moe_puzzle_benchmark_12/
  • frontier_benchmark_30/
  • aeos_orchestration/
  • security_gatekeeper_pipeline/
  • paper_assets/

Security and Open-Source Boundary

• Keep secrets in .env only.
• Do not commit model weights or local checkpoints.
• If you maintain private judge logic for ASRT, expose only integration contracts in public repos.

GitHub Release Checklist

• README.md updated and accurate
• .env ignored
• model weights ignored
• no __pycache__ committed
• verification suite runs: python run_tests.py

License

MIT