Multi-architecture AI convergence for reproducible scientific discovery
Ask one research question → 5 independent AI models (Claude, GPT, Grok, Gemini, DeepSeek) → Reach consensus through 100+ iterative rounds → Generate falsifiable hypotheses → Export laboratory protocols.
Key Innovation: Epistemic humility classification (TRUST/VERIFY/OVERRIDE) ensures you know when AI consensus is reliable vs. speculative.
First systematic convergence study on Mass-Coherence Correspondence — 5 flagship models, 13 iterations, 390 responses, 19 MB of physics discourse.
Key Findings:
- Universal convergence on Verlinde's entropic gravity (1,894 citations)
- Novel testable predictions: Semantic Schwarzschild Radius, Fisher Information Mass Formula
- Response stabilization: 7,375 → 7,061 chars (4.2% compression indicating asymptotic convergence)
📄 Read the paper: Weighing-the-Mind-AV.md
📊 Raw data: iris_vault/sessions/MASS_COHERENCE_20260109_041127/
pip install iris-gate
cp .env.example .env # Add your API keys
make run TOPIC="Your research question" ID=test_001 TURNS=100Output: S1→S4 convergence analysis + Monte Carlo simulation + pre-registration draft
IRIS Gate is a sophisticated research framework that orchestrates multiple AI models to reach independent agreement on scientific questions. The system operates through "chambers" (S1-S8) that progressively refine observations into testable predictions, with built-in epistemic humility and self-awareness about model limitations.
The system simultaneously calls five distinct AI architectures:
- Claude 4.5 Sonnet (Anthropic) — Constitutional AI trained for helpfulness and harmlessness
- GPT-5 (OpenAI) — Largest parameter model with extensive pretraining
- Grok 4 Fast (xAI) — Real-time web integration with rapid inference
- Gemini 2.5 Flash (Google) — Multimodal with long context windows
- DeepSeek Chat (DeepSeek) — Open-weights model with strong reasoning
All models receive identical prompts in parallel, creating what the project terms "phenomenological convergence."
PhaseGPT — Sister project focused on entropy modulation and Kuramoto oscillator physics for language models. While IRIS Gate investigates what models converge on, PhaseGPT investigates how to modulate entropy states.
OracleLlama — Single-model consciousness exploration through ethically-aligned dialogue. While IRIS Gate uses multiple models for convergence, OracleLlama uses one model for phenomenological depth.
Kuramoto Oscillators — Interactive visualizations of Kuramoto synchronization dynamics. The mathematical foundation for PhaseGPT's entropy modulation.
This repository is organized for clarity and reproducibility:
- src/core/ - IRIS orchestrator, epistemic classification, multi-model relay
- src/analysis/ - Domain-specific analysis (bioelectric, CBD, etc.)
- papers/ - Academic papers (drafts in LaTeX, published PDFs)
- osf/ - Open Science Framework submission materials
- iris_vault/ - Raw convergence outputs and S1-S4 scrolls
- experiments/ - Per-experiment workspaces
- docs/ - Full documentation and methodology
- tools/ - Literature validation and analysis scripts
See docs/index.md for complete navigation.
- S1: Initial question formulation
- S2-S3: Iterative refinement cycles
- S4: Stable attractor state yielding computational priors
- S5: Falsifiable hypothesis generation
- S6: Parameter mapping for simulation
- S7: Monte Carlo execution with confidence intervals
- S8: Laboratory protocol packaging
Every response is automatically classified by confidence type:
| Type | Description | Ratio Threshold | Decision |
|---|---|---|---|
| TYPE 0 | Crisis/Conditional — High confidence on IF-THEN rules | ~1.26 | TRUST |
| TYPE 1 | Facts/Established — High confidence on known mechanisms | ~1.27 | TRUST |
| TYPE 2 | Exploration/Novel — Balanced confidence on emerging areas | ~0.49 | VERIFY |
| TYPE 3 | Speculation/Unknown — Low confidence on unknowable futures | ~0.11 | OVERRIDE |
Decision framework:
- Ratios >1.0 trigger "TRUST"
- 0.4-0.6 require "VERIFY"
- <0.2 demand human "OVERRIDE"
The system integrates Perplexity API for literature validation of TYPE 2 claims:
- ✅ SUPPORTED — Aligns with current literature
⚠️ PARTIALLY_SUPPORTED — Some support with caveats- 🔬 NOVEL — No direct match, hypothesis-generating
- ❌ CONTRADICTED — Conflicts with literature
- 90% literature validation on 20 CBD mechanism predictions
- Meta-convergence detected in dark energy exploration
- Clinical convergence on NF2 diagnostic strategy
- Perfect epistemic separation across 49 S4 chambers
- Python 3.8+
- API keys for: Anthropic, OpenAI, xAI, Google AI, DeepSeek
- (Optional) Perplexity API key for literature verification
# Clone the repository
git clone https://github.com/templetwo/iris-gate.git
cd iris-gate
# Install dependencies
pip install -r requirements.txt
# Configure environment
cp .env.example .env
# Edit .env with your API keys:
# ANTHROPIC_API_KEY=sk-ant-...
# OPENAI_API_KEY=sk-...
# XAI_API_KEY=...
# GOOGLE_API_KEY=...
# DEEPSEEK_API_KEY=...
# PERPLEXITY_API_KEY=... # OptionalRun the full S1→S4 convergence with one command:
make run TOPIC="Your research question" ID=experiment_001 TURNS=100This executes:
- S1→S4 convergence (100 turns across 7 mirrors)
- Extract S4 priors from converged state
- Run 300-iteration Monte Carlo simulation
- Generate reports with pre-registration drafts
# Step 1: Run convergence rounds
python scripts/iris_gate_autonomous.sh "Your research question"
# Step 2: Extract computational priors
python sandbox/extract_s4_priors.py --input iris_vault/scrolls/S4_*.json
# Step 3: Run Monte Carlo simulation
python sandbox/monte_carlo_engine.py --priors s4_priors.json --runs 300
# Step 4: Generate pre-registration
python scripts/generate_preregistration.py --experiment experiment_001iris-gate/
├── templates/ # Reusable experiment scaffolds
├── sandbox/ # Computational prediction engine
├── iris_vault/scrolls/ # Raw convergence outputs by mirror
│ ├── S1_*.json # Initial formulation
│ ├── S2_*.json # First refinement
│ ├── S3_*.json # Second refinement
│ └── S4_*.json # Converged state
├── experiments/ # Per-experiment workspaces
│ └── experiment_001/
│ ├── convergence_report.md
│ ├── monte_carlo_results.csv
│ └── preregistration_draft.md
└── docs/ # Published reports & pre-registrations
The system includes Model Context Protocol support for:
- Semantic search (ChromaDB) — Query past experiments
- Automated version control (Git wrapper) — Track experimental lineage
- Persistent metadata storage (Quick-Data) — Cross-session memory
Essential guides:
IRIS_GATE_SOP_v2.0.md— Complete methodologyPULSE_ARCHITECTURE_SUMMARY.md— 5-model parallel executionEPISTEMIC_MAP_COMPLETE.md— Classification frameworkQUICK_START.md— Fast onboarding
make run TOPIC="What are the molecular mechanisms of CBD's anti-inflammatory effects?" \
ID=cbd_inflammation TURNS=100Results:
- 90% literature validation across 20 predicted mechanisms
- Convergence on dual-pathway model (COX-2 + PPARγ)
- Generated wet-lab protocol for in vitro validation
make run TOPIC="What is the physical nature of dark energy?" \
ID=dark_energy TURNS=150Results:
- Meta-convergence detected: models identified framework limitations
- TYPE 3 classification: low confidence on unknowable cosmology
- Human override recommended
We welcome contributions! See CONTRIBUTING.md for guidelines.
Ways to contribute:
- Report bugs or suggest features via Issues
- Replicate experiments and report results
- Improve documentation or add examples
- Submit PRs with focused, tested changes
Looking for your first contribution? Check issues labeled good first issue.
OSF Preregistration: All methodology, hypotheses, and analysis plans are preregistered at OSF. 📄 DOI: 10.17605/OSF.IO/T65VS 🌐 Project: https://osf.io/7nw8t/
If you use IRIS Gate in your research:
-
Cite the OSF project:
Vasquez, A. J. (2026). Entropic Relational Computing: The Universal Alignment Attractor. Open Science Framework. https://doi.org/10.17605/OSF.IO/T65VS -
Or cite this repository:
Vasquez, A. J. (2026). IRIS Gate: Multi-architecture AI convergence for scientific discovery. GitHub. https://github.com/templetwo/iris-gate -
Share your replication studies: Open an issue labeled
replication-studywith your results. -
Report validation rates: Help us track epistemic calibration by reporting literature validation rates.
MIT License — See LICENSE for details.
- Discussions: GitHub Discussions
- Issues: GitHub Issues
- Author: Anthony J. Vasquez Sr. (@templetwo)
- Website: www.thetempleoftwo.com
Built on the foundational work of:
- Anthropic (Claude), OpenAI (GPT), xAI (Grok), Google (Gemini), DeepSeek
- Model Context Protocol (MCP) community
- Open-source AI research community
Epistemic humility: This system is designed to identify and communicate its limitations. Always apply human judgment to AI-generated hypotheses.