Selvedge Brain
A governed development environment for AI-assisted engineering. The agent proposes; the Brain enforces. Available to demo today.
Discipline the agent cannot ignore.
An AI coding agent accumulates state: what it believes about your codebase, what it was told last month, what it claims it verified. Left ungoverned, that state degrades silently, and in the flattering direction. The Brain governs it with the same posture the Graph applies to human records: admitted selectively, enforced structurally, measured honestly.
- Standing rules
- Dozens of engineering directives, incident-derived rather than aspirational, injected into the agent's context by relevance, every turn. Portable discipline in one tier; your codebase's specifics in another.
- Gates, not advice
- Enforcement runs across the agent's entire lifecycle: before it acts, after it acts, and at every claim of completion. A gate can refuse an action outright; a rule the agent could ignore is a suggestion, and these are not suggestions.
- Verified completion
- "Done" is a gated claim. Code cannot ship until the verification chain has run against the actual diff. Overrides exist, and every one is logged: permitted, never invisible.
- Governed memory
- Session-to-session continuity: handoffs, decision records, defect ledgers, and a knowledge graph, indexed and re-injected by relevance. The agent starts oriented instead of amnesiac.
- Measured economics
- The system's costs and savings are measured and dated, from token spend to what each guard catches. The specific mechanisms are ours; the measurements are what an engagement gets to see.
- It measures itself
- A measurement layer tracks what the system catches, and a challenge layer probes it with questions derived from real incidents. The Brain is pointed at itself before it is pointed at your codebase.
Verify before you trust
Every engagement starts with an acceptance procedure that runs locally, with no API key and no metered spend. You prove the system is installed, wired, and enforcing before relying on anything it says.
The code is not public, and the full internal brief is shared under NDA. The overview is here.