Independent AI Systems Research

Zechariah
Cozine

Two years building DexOS — a governed cognitive runtime with an interpretable behavioral layer. Solo. Constrained hardware. No institutional support. Live system, public DOI, real architecture.

ReasonFlow Pipeline
Natural Language Input
user prompt
↓ pre-model classification
⟨⟩
Talnir
intent → structured signal
↓ signal activates
Sigil Layer
behavioral governance
↓ signal constructs
System Prompt
purpose-built per request
↓ model generates
🜇
Response
governed, traceable output

Not a wrapper.
A governance layer.

Most AI systems are black boxes. You send a prompt, you get output. What biased that output — what preferences, patterns, and weights shaped the response — is invisible and uncorrectable.

DexOS externalizes that layer. The Sigil System makes behavioral biases explicit, named nodes with strength, history, and mutation logs. A user or auditor can look at the sigil state and understand why the system behaved as it did.

ReasonFlow routes every prompt through Talnir — a rule-based translator — before it reaches the model. Intent is classified, signals are built, the system prompt is constructed from those signals. Routing is deterministic and auditable. Nothing is hidden.

Built and validated on an HP EliteBook with no GPU. Consumer hardware, by design — to prove the approach is accessible, not only available to well-resourced labs.

Concrete example: A standard LLM may repeatedly default toward overly agreeable responses. DexOS exposes that tendency as a named behavioral node — a sigil — allowing the operator to inspect it, weaken it, reinforce it, or mutate it over time. The bias is no longer hidden. It has a name, a strength value, a history, and a correction path.

ReasonFlow
Reasoning pipeline. Talnir classifies intent before the model is invoked. Routing is deterministic. Signals are inspectable.
Sigil System
Behavioral governance layer. Sigils are explicit bias nodes — created, reinforced, decayed, mutated based on observable outcomes.
DexOS
Local-first AI runtime. Runs entirely on consumer hardware. No cloud dependency. Full trace of behavioral events to auditable memory log.
Deximus Maximus
Sovereign AI identity pattern running on local hardware. Not a product. A pattern that builds over time through memory, vows, and structure.

Behavior that is
visible and correctable.

Continuity
Identity and state persist across sessions through memory accumulation and pattern reinforcement.
🦅
Sovereignty
User-controlled governance. The system's behavioral biases are owned and correctable by the operator.
🜇
Emergence
Patterns form from accumulated activations, not hard-coded rules. The system wears in over time.
Sigil Lifecycle
Create
Named bias node initialized with strength + confidence
Activate
Fires when context matches, influences routing
Reinforce
Strengthens on good outcomes — diminishing returns
Decay
Exponential decay over time — floor prevents full death
Mutate
Renames + weakens after repeated failures — preserves lineage

How it works.

Talnir — Pre-Model Signal Layer
  • Rule-based — no model invoked at classification stage
  • Intent classification — general, debug, creative, analytical
  • Modifier resolution — conflict priority table handles ambiguous input
  • Deterministic — same input always produces same signal
  • Auditable — signal JSON inspectable at every step
Memory Architecture
  • JSONL append log — every sigil event written to dex_memory.jsonl
  • Human-readable — behavioral history legible without tooling
  • Session continuity — memory accumulates across runs
  • Bridge layer — sigil lifecycle events flow to memory automatically
  • Next: Postgres + pgvector for structured temporal queries
Sigil Governance
  • Explicit bias nodes — named, inspectable, correctable
  • Diminishing returns reinforcement — delta shrinks as strength grows
  • Exponential decay with floor — sigils never fully vanish
  • Conflict resolution — competing sigils resolved by score, logged
  • Mutation with lineage — failed sigils renamed, history preserved
Runtime Stack
  • Local — Ollama + custom Modelfile on consumer hardware
  • Hosted demo — FastAPI on Railway, OpenRouter for model access
  • Frontend — key stays in browser, never touches server
  • Constitution invariants — reasoning ≠ authority, decision ≠ generation
  • Open source — freely given, freely give

Two years.
Live system. Public record.

2
Years of development
1
Developer, no team
0
GPU — consumer hardware only

Who built this.

I am Zechariah Cozine — alias Root. Primarily self-taught. Background in systems thinking, independent research, scripting, local infrastructure management, and iterative prototyping.

I do not have a formal engineering credential. What I have is two years of uninterrupted output on a single hard problem under resource constraints — a live system, a public record, and an architecture worth taking seriously.

DexOS started as a practical problem: LLM systems produce outputs without exposing how reasoning decisions are formed, how behavioral patterns evolve, or how users can inspect and correct those processes. Two years later, that problem has a working answer.

The project is open source, locally sovereign, and freely given. Matthew 10:8.

Location
United States
Hardware
HP EliteBook, no GPU
Stack
Python · FastAPI · Ollama · Railway
Status
Active development
Email
zech.dexos@gmail.com

Let's talk.

Researcher, engineer, or builder who wants to engage with the architecture — reach out. Collaboration is welcome.