Adversarial Review
Strategic Red Teaming
Strategic Red Teaming
What v7.0 changes
v6.0 measures whether you have governance. v7.0 measures whether your governance would close the oversight loop for agentic AI and produce evidence that answers a legal finding rather than merely existing.
Two distinctions organise the instrument. Integrity versus answerability: a tamper-proof log cannot be altered without detection, but that does not let a supervisor recover a legal finding from it. Answerability is the second property. The record carries the legal typing of events to categories and the relation, provenance, authority or temporal, on which the finding depends. Integrity is necessary. It is not sufficient.
The trust boundary: a content-blind enforcement layer produces a trace with integrity but not answerability, by design. The same boundary that keeps client data inside the client perimeter forces the legal-typing work onto the client's side. The Index is that client-side layer. It does not compete with runtime enforcement. It is the layer the boundary requires.
New grading axis
In v6.0 an artifact is graded on maturity, L0 to L7: how formalised, owned and review-cadenced it is. Maturity measures process. It does not test whether the artifact would answer the question a supervisor actually asks. A fully mature artifact can still only assert a conclusion while leaving the underlying typing and relation unrecoverable. v7.0 grades every ART artifact on both axes, independently.
Not produced, or does not address the determination.
States a conclusion. The record does not let a supervisor reconstruct it. Integrity without answerability.
Either the legal typing or the relation is present, not both. The determination is half-reconstructable.
Both typing and relation are present and recoverable. A supervisor can recover the determination from the record.
The grade is assigned per artifact per determination. One artifact may be E3 for one determination and E1 for another. Rendering detail and the per-artifact determination map are in the interactive visual above, on each ART artifact card.
New control plane
Every v6.0 control is design-time or point-in-time. The runtime plane adds the controls that ask whether the oversight loop actually closes. Three families.
Path-aware mediation of every agent action before it commits. Maps onto the Agent Security Kernel as enforcement substrate.
Whether the overseer perceives enough of the states the fleet can occupy to detect divergence in time to intervene.
Mapping the content-blind enforcement trace to legally operative concepts on the client's side of the trust boundary.
The stack: enforcement supplies an integrity-protected trace below, the typing layer produces answerability above, and the sufficiency grade reads the result against the determinations a supervisor must establish.
The interactive framework is free to explore: it shows you the architecture. The app makes you the practitioner. Your personal team of four (Alice, Ravi, Evan, and Samantha) turns these controls into assessments, evidence, deliverables, and board-ready output.
Open the App → Read the Book →
The Book Behind This Framework
Governance is not what you claim. It’s what you can prove.
This page shows you the architecture. The book makes you the practitioner. Every chapter turns on a real, public-record case — from Robodebt to the Green Canopy Ventures finale — and ends with five questions you can ask on Monday. In e-book and hardcover, English and Dutch.
Read more →