Scientific Matrix

An experimental architecture for AI-assisted research, engineering, memory and verification.

Status: early-stage research prototype, published as a preview. This repository communicates the vision, the high-level design and one tiny educational demo. It is not the full system, is not production-ready, and deliberately does not contain the working implementation.

---

Modern AI systems are increasingly capable of writing code, using tools and performing research. The harder problem is coordinating those capabilities reliably over long-running work — so that an agent doesn't just sound confident, but actually produces work that can be checked.

Scientific Matrix explores an architecture where specialized components can:

• plan — decompose a goal into tractable steps
• research — gather and summarize relevant material
• write code — produce candidate artifacts
• use tools — call deterministic programs instead of guessing
• remember — carry useful results forward across experiments
• simulate — test ideas in cheap sandboxed environments
• verify — check outputs independently of how they were generated
• iterate — feed verified results back into the next cycle

…while keeping evidence and verification separate from generation.

The idea in one line

GENERATE ≠ VERIFY — the component that produces a result must never be the component that certifies it.

The loop, conceptually

        Goal
          ↓
       Planner
          ↓
  Specialized agents
          ↓
 Tools / Code / Simulation
          ↓
 Independent verification
          ↓
         Memory
          ↓
    Next iteration

Each pass through the loop produces artifacts (code, findings, models of what was tried) and verdicts (independent checks on those artifacts). Only verified material is allowed to become durable memory.

Repository layout

README.md               ← you are here
LICENSE
SECURITY.md             ← how to report issues, and the limits of this preview
CONTRIBUTING.md         ← contribution philosophy for a preview repository
docs/
  ARCHITECTURE.md       ← high-level architecture, component responsibilities
  PRINCIPLES.md         ← the design principles behind the loop
  ROADMAP.md            ← broad stages, current status, honest limitations
examples/
  minimal_matrix_demo/  ← one tiny, offline, stdlib-only Python demo of the loop

Relationship to Canary

Scientific Matrix and Canary are separate projects.

• Matrix explores autonomous research and engineering workflows.
• Canary is an independent verification layer for AI-written software.

Conceptually:

Matrix builds.   Canary verifies.

They are complementary by design — generation and verification belong to different systems — but no production integration between them is claimed or implied at this stage.

What exists here vs. what doesn't

	In this repository	Not in this repository
Vision, principles, architecture	✅ documented	—
Working orchestration engine	❌ described only	✅ lives privately, in development
Memory / verification internals	❌ interfaces only	✅ withheld
Runnable code	only examples/minimal_matrix_demo/	—

This split is intentional: the public repository shows the interface and the idea; the recipe stays private while the system matures.

Quick start (the demo)

cd examples/minimal_matrix_demo
python matrix_demo.py            # honest worker: results get verified
python matrix_demo.py --sabotage # worker plants a wrong answer: verifier catches it

No dependencies, no network access, fully deterministic.

Disclaimer

This is a research exploration, not a product. Expect rough edges, missing pieces, and language that says experimental far more often than done.

License

Released under the MIT License. The demo and documentation in this repository are free to use; the private implementation they describe is not included and is not licensed here.