TO2D

Architecture Lab

2026–

The Deterministic Shell Architecture

Modern AI systems are stochastic at their core. Production systems cannot be. This stack isolates randomness, captures execution signals, and enforces admissible state transitions so automation remains stable under uncertainty.

raw outputLLMSTOCHASTIC COREvalidated actionllm-contractINVARIANT BOUNDARYdetectsschema violationshallucinated fieldstype mismatchesstructural driftexecution signalsBrowserAgentOPERATORdetectsselector failurestimeoutauth rejectionunexpected navigationBrowserStateSIGNAL CAPTUREdetectssession corruptionstate divergencemissing evidenceunreplayable conditionsinformation flow ↓

BrowserAgent — Operator Layer

Executes actions in an unstable environment. All world-facing signals originate here.

  • UI interaction
  • Network navigation
  • Auth flow
  • Tool calls

BrowserState — Signal Capture Layer

Captures execution state as replayable signals. Makes failures observable instead of anecdotal.

  • Cookies / storage
  • Navigation graph
  • Error evidence
  • Reproducible state snapshots

llm-contract — Invariant Boundary

Enforces admissible output states around stochastic model behavior.

  • Schema validation
  • Invariant enforcement
  • Categorized failure modes
  • Targeted repair loop

Stochastic Core, Deterministic Shell

The model remains probabilistic.

The boundary is deterministic.

Execution remains adversarial.

State becomes measurable.

This is the architecture.

Chapters

System Identification

Characterizing LLM behavior as dynamical systems. Inputs, outputs, transfer properties.

Read full chapter →

Stochastic Core

Modeling the irreducible randomness in language model outputs. Distributions, not determinism.

Read full chapter →

Invariant Boundary

Structural constraints that hold regardless of model, prompt, or context window.

Read full chapter →

Zero-Context Architecture

Domain isolation, minimal representations, and reproducible agent behavior.

Read full chapter →

Fuzzy Systems

Modeling decision systems under uncertainty through membership activation, rule inference, and output control.

Read full chapter →