Gauntlet 🥊

A combinatorial multi-agent stress test for high-stakes research proposals.

Gauntlet is a "virtual brainstorming engine" designed to subject research proposals (specifically one's that are a bit cross-cutting across areas) to intense scrutiny from simulated domain experts. It uses the Anthropic Claude Opus 4.5 API to create a divergence-convergence feedback loop.

Instead of a single review, Gauntlet explodes the solution space:

Divergence: Three distinct expert personas each critique the proposal three times at different temperatures (0.3, 0.7, 1.0), naturally producing conservative, balanced, and divergent takes (Total: 9 distinct critiques).

Convergence (The Flywheel): A "Synthesizer" agent takes every combination of one critique per expert (3³ = 27) and generates a unique strategic synthesis report for each, helping you find the "Golden Thread" narrative that satisfies all constraints.

Names below are fictitious names to just explain the readme!

🗂 Project Structure

Gauntlet/
├── inputs/
│   ├── proposal_call.pdf               # The funding solicitation
│   └── my_proposal.pdf                 # The preliminary draft
├── personas/
│   ├── dr_silas_vane.md                # Chaos Control & Swarm Dynamics Expert
│   ├── prof_amara_kito.md              # Bio-Photonic Interfaces Expert
│   ├── dr_julian_rex.md                # Exascale Cognitive Systems Expert
│   └── synthesizer.md                  # The Strategy Lead System Prompt
├── outputs/
│   ├── RUN_CONFIG.md                   # Run metadata & naming guide
│   ├── expert_reviews/                 # The 9 critiques (3 personas × 3 temps)
│   │   ├── dr_silas_vane/              #   run_1.md  run_2.md  run_3.md
│   │   ├── prof_amara_kito/
│   │   └── dr_julian_rex/
│   └── syntheses/                      # 27 self-contained action-plan folders
│       ├── silas_1__amara_1__julian_1/ #   SYNTHESIS.md + 3 source reviews
│       └── …
├── generate_persona.py                 # Optional: auto-generates persona .md files
├── .env                                # ANTHROPIC_API_KEY
├── config.toml                         # Persona & synthesizer selection
└── main.py                             # Orchestration & permutation logic

🧠 The Expert Panel (Simulated)

Dr. Silas Vane: Expert in Chaos Control & Swarm Dynamics.

    Role: Focuses on emergent behavior, robustness, and mathematical proofs of stability in unpredictable environments.

Prof. Amara Kito: Expert in Bio-Photonic Interfaces.

    Role: Focuses on the physical layer, energy constraints, biocompatibility, and the "wetware" interface constraints.

Dr. Julian Rex: Expert in Exascale Cognitive Systems.

    Role: Focuses on massive scalability, foundation model architectures, and distributed system latency.

🚀 The Logic Flow

Ingestion: The engine reads the proposal_call.pdf and my_proposal.pdf.

The Gauntlet (Phase 1 — Divergence): Each expert reviews the proposal three times, once at each temperature (0.3, 0.7, 1.0). Low temperature yields precise, conservative critiques; high temperature yields creative, divergent pivots. This produces 9 distinct critique vectors with no manual prompt engineering.

The Flywheel (Phase 2 — Convergence): The system computes the Cartesian product — one run per expert — giving 3³ = 27 unique input combinations. The Synthesizer agent produces a cohesive Action Plan for each. Every output folder is self-contained: SYNTHESIS.md plus copies of the three source reviews that fed it.

Idempotent Resume: If you hit a rate limit mid-run, just re-run main.py. It skips any output that already exists and picks up exactly where it left off.

💡 The Idea Generator (Optional Front-End)

Before running the Gauntlet you can use idea_generator.py to produce the proposal kernel itself. It takes a baseline paper and a rough seed idea, then drafts a proposal that has already been written against the same expert personas — anticipating their critiques before they happen. The output is both a Markdown file and an auto-generated PDF. You edit the markdown if needed, then feed the PDF into the Gauntlet as your proposal. See IDEAGEN.md for the full concept and how-to.

🛠 Usage

Install dependencies:
Bash

pip install anthropic pypdf python-dotenv markdown-pdf

Note: `markdown-pdf` is only needed for the Idea Generator's automatic PDF conversion.

Add your API keys to .env:
Plaintext

ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...           # Only needed for idea_generator.py

Edit config.toml to select which personas and synthesizer to use.

Run the Gauntlet, passing your two PDFs as arguments:
Bash

python main.py inputs/proposal_call.pdf inputs/my_proposal.pdf

Use -o to write output to a custom directory:
Bash

python main.py -o runs/nsf_2026 inputs/proposal_call.pdf inputs/my_proposal.pdf

The source code by default has TEMPERATURES: list[float] = [0.3]. Just comment that out and uncomment the previous line to get the full-blown 3 critiques per expert.

🧬 Generator Tool (Optional)

If you don't want to write personas by hand, use the generate_persona.py script. It scrapes a researcher's website and uses Google Gemini to "clone" their writing style and expertise into a .md persona file in personas/.

1. Get a Google API Key (Free tier available): Google AI Studio

2. Add to .env:

   GOOGLE_API_KEY=your_key_here

3. Install requirements:

   pip install google-generativeai beautifulsoup4 requests

4. Run the generator:

   python generate_persona.py

   Input: "Geoffrey Hinton" Input: "Deep Learning" Input: https://en.wikipedia.org/wiki/University_of_Toronto

   Result: personas/geoffrey_hinton.txt is created automatically.

🎛️ Tuning the Synthesizer The personas/synthesizer.md file acts as the "triage" of this brainstorming. You should edit this prompt to match the specific funding opportunity you are targeting. You must also instruct the synthesizer on your specific goals: do you want it to find a cohesive narrative (the "Golden Thread"), conduct a brutal risk assessment, or focus on commercial viability? Tailoring this prompt ensures the final Action Plan is optimized for your specific success criteria.

📄 License

MIT