Step 1. What the category is
A prompt that drives an autonomous system decides what that system does. That makes it a specification, and every other artefact with that property is versioned, reviewed, addressed and audited.
Most prompts are none of those things. They live in a string, a database row or a source file, and the more careful case, the source file, still only tells you what the file said at a commit. It does not tell you what the model was actually given at execution, after templating, after retrieval, after composition with other units. The artefact that ran is not the artefact that was reviewed.
Verifiable Agentic Prompt Architecture is the name for closing that gap. Four guarantees, and none of them stands up alone:
| Guarantee | Question | Without it |
|---|---|---|
| Typed contract | What shape is this unit, and how can it fail | Failures are found in production, one at a time |
| Identity | Is what ran what was approved | Review is theatre |
| Blast radius | What else breaks if this changes | Every edit is a coin flip |
| Record | What happened, and did quality move | Improvement is indistinguishable from regression |
GRACE is the protocol. STRATT is the reference implementation. This site is where the category is explained and where the claims about it are kept checkable.
Why it is not just careful software engineering
Section titled “Why it is not just careful software engineering”That is the strongest objection and it is half right. Versioning, review, dependency analysis and audit logging are ordinary. The object is not: a prompt has no compiler to reject it, no build-time type error, and no test that fails deterministically. Software practice assumes an artefact that fails loudly. A prompt fails plausibly, and every guarantee above exists to reintroduce a failure signal the substrate does not provide.
The long form of this argument, including what would prove it wrong, is Why Verifiable Agentic Prompt Architecture is its own category.
Step 2. What the protocol defines