Is this real, or another AI governance metaphor?

“Isn't this just Zero Trust with a new coat of paint?”

No. Zero Trust verifies who you are and whether you can access something. KTP verifies what you are about to do, on what object, at what speed, with what blast radius. Zero Trust controls access. KTP controls consequence. KTP sits on top of Zero Trust, not next to it.

“Isn't this ABAC / PBAC / IAM / SOAR / DLP with new vocabulary?”

Those control identities, attributes, policies, response workflows, and data movement. KTP composes them around a unit they were not built for: an attempted action by an autonomous agent under context. The plumbing is familiar. The unit of governance is new.

“Not all AI risk is action-based. What about bias, labor displacement, training data?”

Correct. Most of those risks are not motion risks. The MIT scoring puts roughly half the risk universe outside KTP's direct scope. KTP governs the action layer. Other layers need other instruments. We say so on every page.

“Verbs alone are too simplistic. Read by itself isn't a control.”

Also correct. The unit of evaluation is not the verb. It is the tuple: agent, verb, object, channel, context, consequence, constraint. Verbs expose the control surface. Context determines the constraint. The verb library is the index, not the policy.

“How does this work outside a controlled enterprise environment?”

It mostly doesn't, yet. KTP assumes a runtime where actions can be intercepted at a tool gateway, API, identity layer, or egress point. On the open internet, against unmanaged agents, the enforcement surface shrinks to whatever your edge controls. Honest answer: enterprise first, federation later.

“What does 'governable' actually mean? That word is doing a lot of work.”

It is. We define it as a ladder: not visible, observable, attributable, constrainable, interruptible, reversible, pre-authorized. KTP doesn't promise that every risk becomes preventable. It claims to move risks up that ladder. The governability ladder lives at methods; see §03 here for the term sheet.

“Is KTP a protocol, a model, a policy framework, or a product?”

It is a control methodology with a reference architecture and a small vocabulary. Not a product. Not a model. Not a single protocol. The artifacts are: a decision function, an action tuple, a verb library, a set of enforcement surfaces, a governability ladder. Anyone can implement it. Nobody owns it.

“Where is the empirical evidence?”

Thin, on purpose. The v0.2 MIT Scorecard is a deterministic first-pass review expansion, not human adjudication or empirical validation. The 337 generated test cases have not been run against a prototype. We say this on /enterprise/methods. v0.3 is where evidence starts. Today is positioning and architecture.

“The should gate is just un-instrumented telemetry. Wire up an HR feed and 'there was a layoff' becomes a signal like any other.”

Half right. Instrumentation gives you the fact, layoff = true. It never gives you the mapping, therefore the cheerful newsletter is cruel. Stability signals are the constraint directly; the layoff signal is not, and turning a fact into fittingness needs a model that holds norms. Fully instrumented, you still need that model. That residual is exactly why the normative layer is a separate thing, not a SIEM feature. And yes, the should gate is earlier-stage than the can gate. We say so.

• KTP solves AI risk.
• KTP is the answer.
• KTP governs the universe of AI risk.
• KTP makes AI safe.
• KTP is mind-blowing.

• a control methodology
• a bounded contribution
• the action layer of AI risk
• technically governable subset
• enforcement surface
• translation layer between taxonomy and control