Crosswalks & Controls
One control structure, mapped to every rule that matters — feeding both the assessment and the engine.
Confidential · June 2026. Prepared for the govrn.ai data room. Any figures herein are illustrative placeholders, not results.
govrn assesses a company through three lenses and proves each finding against the exact external rule it satisfies. The mechanism that makes this defensible — and reusable — is the crosswalk library: a single, maintained mapping from the govrn control structure to the published frameworks regulators, boards, and customers already recognize. The assessment cites it. The engine consumes it. That shared dependency is the differentiator.
The Three-Lens Model
A company's AI exposure does not live in one place, so the govrn standard reads it across three lenses:
- Technology — how AI systems are built, deployed, and maintained: architecture, data handling, lifecycle controls.
- Security — adversarial surface: prompt injection, model abuse, supply-chain and access risk.
- AI — the model behavior itself, including the product AI a company ships to its own customers — the layer most assessments skip.
The lenses are deliberately overlapping. A single finding — say, an unmonitored model endpoint — surfaces under Technology (lifecycle), Security (exposed surface), and AI (behavioral drift). Reading it three ways is how the assessment avoids blind spots that a single-axis checklist creates.
The Crosswalk Concept
A crosswalk maps every deliverable to the exact rule it satisfies. Not "aligned with NIST" in the abstract — this control answers this clause of this framework. The govrn standard maintains a crosswalk library across six published frameworks:
| govrn lens | Control intent | Framework references |
|---|---|---|
| Technology | AI lifecycle governance, mapping & measurement | NIST AI RMF · ISO/IEC 42001 |
| Technology | Organizational management system, accountability | ISO/IEC 42001 · CSF 2.0 |
| Security | Application & LLM threat controls | OWASP · MITRE ATLAS |
| Security | Adversarial ML & abuse tactics | MITRE ATLAS · CSF 2.0 |
| AI | High-risk obligations, transparency, conformity | EU AI Act · NIST AI RMF |
| AI | Product-AI behavioral review | NIST AI RMF · ISO/IEC 42001 |
Because the mapping is explicit, a govrn finding is traceable in both directions: a board can ask "which regulation does this close?" and a regulator can ask "where is the evidence?" — and the same crosswalk answers both. Attestation is always independent: the body that builds is never the body that attests, and govrn is never self-certified.
Why One Library Feeding Two Surfaces Is the Moat
Most governance tooling builds its mappings twice — one set for the consultant's assessment, another buried in the platform. They drift. govrn maintains a single crosswalk library that feeds both surfaces:
- The govrn standard (the assessment — See + Prove) uses it to map findings to rules for the CISO and the board.
- The Aperture engine (Run) consumes the same control structure to attribute what it measures back to the rules those measurements satisfy.
When a framework updates — a new EU AI Act guidance, a CSF 2.0 revision — the library is corrected once and both surfaces inherit the change. The assessment and the engine never disagree about what a control means, because they read from the same source. As the engine accumulates an attributed, time-stamped record, every measurement is already pre-mapped to the rule it evidences. That is the compounding asset: a crosswalk library that turns raw measurement into attestable proof without re-mapping it.
A note on status, stated plainly. The govrn standard and its six-framework crosswalks are live today. The Aperture engine is built and tested — deterministic, with no model in the measurement path — but pre-MVP on live data; its connectors have not yet run against a live API. The crosswalk library is what guarantees that, once connected, what the engine measures and what the standard proves speak the same language from day one.
See the team building the engine and the roadmap for the connector and isolation milestones ahead.
Two controls the platform vendors validated
The major agent platforms are converging on patterns govrn encodes as vendor-agnostic controls — with those platforms as accepted reference implementations, never as a requirement to use them. Encoding them makes govrn the standard the platforms themselves map to.
- Delegated-token control (identity & access). Agents must not inherit long-lived, human, god-mode credentials. Access must be scoped, short-lived, delegated, and traceable (user → agent → service). Vercel Connect's token model is one accepted implementation; the control is satisfied by any equivalent.
- Autonomous-agent control (agent operations & HITL). Autonomous agents must operate plan-first with batch-scoped approval, execute in isolation, and gate any production-state change on human approval. Vercel Agent's plan-mode + batch-permission-grant + sandbox-before-prod is one accepted implementation.
Both controls crosswalk to NIST AI RMF (Govern / Manage), ISO/IEC 42001, and the agent-security guidance in OWASP / ATLAS. govrn requires the outcome; the platform is one way to get there.