MCOP

A deterministic, replayable trust substrate for AI agents; Merkle-chained provenance for every reasoning step!

Comment

Inspiration

Most "reasoning" claims from AI agent frameworks can't be replayed or audited.
Retrieval-augmented and chain-of-thought wrappers produce outputs you have to
trust on faith. MCOP was built on a different principle — falsify first:
rerun the deterministic benchmark harness before trusting the concept layer.
We were also inspired by stigmergy in biological systems (e.g. ant colonies),
where agents coordinate through persistent traces rather than direct messaging.

What it does

MCOP Framework 2.0 is a recursive meta-cognitive optimization protocol for AI
agents — a deterministic reasoning pipeline where every step is replayable,
byte-identically reproducible across Node, browser, and edge runtimes, and
auditable through a Merkle-chained provenance trail. At its core is a
Deterministic Triad:

  • NOVA-NEO Encoder — a SHA-256 hashing pipeline producing fixed-dimension
    context tensors with entropy estimates.
  • Stigmergy v5 — a vector "pheromone" memory using cosine resonance scoring
    and Merkle-proof chaining.
  • Holographic Etch — a confidence-gated, append-only ledger of state changes
    and eudaimonic scores.

Instead of optimizing for raw throughput alone, MCOP rewards flourishing
trajectories (high alignment + high utility) via eudaimonic scoring.

How MCOP was built

The project is a pnpm monorepo with a Next.js 15 / React 19 application, a
publishable TypeScript library (@kullailabs/mcop-core), and a Python sibling
SDK (mcop on PyPI) that maintains cross-language parity. Determinism is
enforced with RFC 8785 JSON Canonicalization so that
$$\text{SHA-256}(\text{RFC-8785-canonical-JSON}(\text{payload}))$$
is byte-identical across Node crypto, pure-JS SHA-256, WebCrypto subtle, and
Python hashlib. Later layers add a 150-node Proteome substrate, an opt-in
ThermoTruth free-energy layer ( F = U - T \cdot S ), a Drift Sentinel Kernel,
CUDA/ONNX hardware acceleration, and an adapter mesh for Grok, Claude, DeepSeek,
Kimi, Qwen, Magnific, Utopai, and generic REST/MCP/HTTP providers.

Challenges ran into

The hardest part was crystalline determinism — guaranteeing byte-identical
output across runtimes and languages, which required canonical encoding, golden
fixtures, and deterministic zero-padding of ragged vectors. Bounding memory
pressure while keeping the pheromone substrate homeostatic (adaptive resonance
thresholds, circular-buffer eviction) was another challenge.

What was learned

Reproducibility and provenance are features, not afterthoughts. Making every
claim replayable ("falsify first") forced a discipline of kernel-derived,
Merkle-sealed metrics that are never hand-written.

What's next

Distributed cluster mode, hosted provenance ledger, external benchmark replay
recruitment, and deeper hardware acceleration.

Built With

Share this project:

Updates

Submission history