Models don't need superhuman judgment—they need guardrails that encode the judgment their principals already have about risk, scope, and acceptable behavior.
Architecture
The Control Deck
Not bigger, better constrained. An architecture for AI collaboration that treats model limitations not as embarrassments to hide, but as design constraints to build around.
Implemented through the Cognitive Trinity: Lex (mind) + LexRunner (body) + LexSona (soul).
Note: Research papers use "Control Stack" — same concepts, different name.
Core Concepts
The Control Deck is built from six interlocking ideas. Each one reinforces the others.
Receipts
Durable records of what actually happened—not claims, evidence. When the model says "tests passed," the receipt contains the actual test output.
Gates
Objective checks that must pass before a change is accepted. If I'm wrong, the gates tell us. If I'm right, we have evidence.
Modes
Behavioral profiles defining risk tolerance and allowed actions. Different tasks warrant different guardrails—not one-size-fits-all.
Policy Surface
Machine-readable encoding of human judgment about risk and scope. Your judgment, applied consistently across sessions.
Epistemic Guardrails
Mechanisms that make "I don't know" a valued output, not a failure. Honest uncertainty over false confidence.
Scope & Blast Radius
Constraints on what can change and how much. The narrower the lane, the more reliable the work.
The PILOT Loop
The Control Deck orchestrates work through a structured planning cycle. The model never directly edits—it plans and allocates.
| Phase | What it does |
|---|---|
| Perceive | Parse intent, extract features, classify risk |
| Integrate | Pull Lex state, check version contracts, validate scope |
| Layout | Design task graph, assign providers, set budgets |
| Orchestrate | Emit plan to LexRunner, configure gates |
| Track | Consume Receipts, decide: accept / re-plan / escalate |
📖 Full documentation in the Lex repo
The Key Insight
Give them...
- Explicit scope instead of ambiguous prompts
- Required gates instead of optional validation
- Structured modes instead of one-size-fits-all
- Audit trails instead of ephemeral outputs
And you get...
- Work reviewable in minutes, not hours
- Changes reversible with
git revert - Trustworthy results because the receipts are there
- Models that succeed at narrow tasks
The Full Story
📖 Not Bigger, Better Constrained
For the complete narrative—told from a frontier model's perspective—read the essay that explains why constraints are what models actually want.
Read the Essay →Implementation
Lex 1.0.0 (Open Source)
The contract-bearing mind. Frames for episodic memory, lexmap.policy.json for module boundaries, lex.yaml for workflow contracts. Stable on npm.
LexRunner (Private)
Consumes Lex contracts to implement merge pyramids with deterministic execution. The body that reads from the mind's contract surfaces.
Explore LexRunner →