Context Engineering Research

Structured context
is a world model.

A postcode-addressed knowledge base satisfies the formal requirements of a world model: state representation, action-conditioned prediction, and planning support. Independent researchers arrived at the same conclusion from different directions. This is the evidence.

440K
Lines of working code
across 7 repositories
0%
Tier-0 safety violations
down from 20%
100%
Adversarial defense
20 of 20 attacks
14 mo
Building the same insight
8 architectures, one seed

Thesis

P(y|x,C) → certainty as constraint quality increases H(Y|x,C) → 0 semantic entropy collapses under constraints Empirically verified: Tier-0 violations 20% → 0% constraint hardening Adversarial defense 100% 20/20 attacks defended Cost per prevention $0.009 11,234x ROI Self-compilation F(F) ~ F the compiler compiles itself

The formal thesis formalizes context engineering as linguistic infrastructure for AI systems. Semantic entropy, functional determinism, and a 7-type constraint taxonomy with tolerance derivation from incident budgets.

An 80-paper governance audit found that prompt-level constitutional rules produce zero statistically significant improvement (p < 10-14). Structural enforcement works. Advisory rules don't. Validated independently by SimuRA (CMU), Meta FAIR's Code World Model, and the "From Word to World" paper showing 7B models achieve 99.87% accuracy as world models for structured domains.

Research

ML.01

The Formal Thesis

Context engineering as linguistic infrastructure. P(y|x,C), semantic entropy, functional determinism, constraint taxonomy.

ML.03

GEOMETRY Spec

14 constraints for multi-agent compilation. First-principles formal specification with 4 composition constructors.

ML.05

SEED Protocol

Four Laws, 5 rejection patterns, 200-property GENOME across 6 constitutional layers.

ML.07

Two Forces

Expansion and collapse as universal pattern. The compiler as a progressive forbidding machine.

ML.09

Agentic City

Multi-agent operating system. Token, Agent, Cluster, Department, Agency, City, Civilization.

ML.15

Asymmetric Dialectic

Entity vs Process lens. 3-turn debate per pipeline stage. The generative mechanism inside compilation.

RS.17

Polycentric Governance

PCG architecture from 165 sources. Ostrom, Arrow, BFT. Five-layer synthesis for multi-agent systems.

RS.26

The Gap

Structured context is a world model. SimuRA validates the representation. Nobody has published this connection.

ML.14

The Genesis

From tattoo consultations to semantic compilation. The origin story in the builder's own words.

ML.06

Convergent Guarantees

Two independent approaches arrived at the same architecture. The structure is a theorem, not a design choice.

RS.28

80-Paper Audit

Prompt-level rules: zero improvement. Structural enforcement: proven. The empirical basis for constraint-first design.

ML.17

Seed-Builder CLI

Working blueprint compiler. Multi-provider. DDC cross-compilation. Actual self-compilation outputs.

58 postcoded articles across 4 domains. Every claim traces to source. Full wiki on GitHub

Systems

Compiler

ada-seed-engine

Compiles natural language intent into governed multi-agent topologies. 14 structural constraints. Emits OpenClaw workspaces.

TypeScript · 4,011 LOC · 2,240 tests
Training

context-world-model

Training pipeline for the first context world model. LoRA fine-tune on Qwen2.5-1.5B. NVIDIA DataDesigner integration.

Python · Pipeline · A100-ready
Foundation

motherlabs

200 genome properties, 7-agent pipeline, perception, social publishing, self-improvement daemon.

Python · 126,000 LOC · 9,695 tests
Production

Semantic Compiler

Web workbench, async compilation API, governance reports, export bundles for downstream coding agents.

Python + Next.js · 246,998 LOC · 7,079 tests
Knowledge

alex-wiki

58 postcoded articles. Provenance-tracked. Multi-agent access. The world model in action.

Markdown · 4 domains · INGEST/QUERY/LINT
Biology

abiogenesis

Self-replicating programs from random byte soups. No fitness function. No design. Just physics.

Python · pip-installable · MIT
DEPENDENCY MAP motherlabs (Python, 126K) ---- SEED, GENOME, AXIOMS | v Semantic-Compiler (247K) ---- PRODUCTION DEPLOYMENT | v ada-seed-engine (4K) ---- GEOMETRY.md, 14 CONSTRAINTS | | v v OpenClaw workspaces context-world-model ---- TRAINING | v Agents in production

github.com/alexrozex

Timeline

Feb 2025
First concept. A hand-drawn tree. "I give AI what I imagine to be a finished project, and it takes it apart."
Jul 2025
Formalized as recursive cognitive scaffolding. The ambiguity gate appears: stop and ask, don't guess.
Oct 2025
"I figured out what Motherlabs is." The CEO/Chief-of-Staff model. Context engineering platform, not a chatbot.
Nov 2025
Mathematical proof of self-improving software by construction. L0-L5: 93.8% convergence. L7: 7/7 autonomous.
Dec 2025
Constraint-hardened engine. Safety violations: 20% to 0%. Adversarial defense: 100%. Cost: $0.009 per prevention.
Jan 2026
13 iterations across Cursor, Claude, Gemini, Codex, Kimi. MOTHER daemon runs 30+ self-improvement cycles overnight.
Feb 2026
Two Forces treatise. Convergent Guarantees. "The architecture is a theorem, not a design choice."
Mar 2026
GEOMETRY.md: 14 constraints. First-principles formal specification. Self-compilation test passes.
Apr 2026
58-article knowledge base. Context world model training pipeline. 6 mining agents recover 440K lines from archives.

About

Name
Alex Roze
Age
34
Location
Vancouver, BC
Background
Tattoo artist, 15 years
Education
Self-taught
Building since
2024
GitHub
alexrozex
X
@alexrozex

I'm a tattoo artist who independently discovered concepts from compiler theory, formal verification, and cognitive science. Then spent 14 months and $10,000 building the system.

When a client says "wolves, nature, not cheesy," I don't ask twenty questions. I excavate the image they can't articulate. I've been doing semantic compilation on skin my whole career. I just didn't know the name for it.

I don't read code. I don't write code directly. I operate at the intent layer, describing what should exist and working with AI to make it real. The 440,000 lines of working code across seven repositories were built this way. That's not a limitation. That's the proof that the system works.

The formal thesis was validated independently by researchers at CMU, Meta FAIR, and multiple papers on ArXiv. None of them knew this work existed. When independent approaches converge on the same structure, the structure is probably real.

Previously: Rocky Mountain Tattoo, 4 locations, Canada · Princelet Tattoo, London · Latvia → London, 13 years → Vancouver