Prompt-only governance
Important constraints should live in inspectable files and receipts, not only in system prompt text.
Concept reference
The Control Deck is the vocabulary around bounded autonomy.
These concepts explain how agents can act with room to reason while still producing reviewable work: receipts, gates, modes, policy surfaces, epistemic guardrails, and scope.
Core concepts
Receipts
Structured evidence of what was attempted, decided, blocked, or completed.
Gates
Checks that decide whether work can advance, must stop, or needs escalation.
Modes
Operating posture for the current job: exploration, implementation, review, or operation.
Policy Surface
Machine-readable boundaries that help agents reason before they act.
Epistemic Guardrails
Rules that make uncertainty and missing context explicit instead of hidden.
Scope and Blast Radius
How much state, authority, and surface area a run is allowed to affect.
Where they appear in the stack
| Lex | Stores Frames, recalls context, and records receipts for work state. |
|---|---|
| AXF | Declares capabilities and can attach metadata about expected effects and safety. |
| LexThority | Evaluates authority envelopes and writes allow, escalate, or deny receipts. |
| LexSona | Returns behavioral constraints that shape the mode and posture of a run. |
| LexRunner | Coordinates flow through gates, contained execution roots, and promotion decisions. |
What this avoids
Rigid SOP automation
Agents should still reason. The deck bounds effects and records decisions without reducing every run to a brittle script.
Unbounded tool access
Tools and credentials stay governed by existing systems; the stack makes the agent's intent and failures clearer.