Does the spec confidence index predict output confidence, and over what horizon?
Section 4 · Research blueprint
A5 — Spec-Driven Assurance, end to end
An executable, quantified specification is the primary artifact. Code, tests, evaluation sets, and monitors are generated from it and traced back to it.
A5 is primary on the Intent Layer and secondary on every other layer — the first anchor to touch all five. It closes three open threads at once: specification-by-example as the primary artifact, evidence-budget allocation, and confidence rubrics by anchor.
The distinctive claim. A2 showed that hardening intent narrows the space of plausibly correct implementations. A5 makes the same hardened intent generate the evidence plan, so the verifier is anchored to the spec rather than to the code the generator happened to write. The Two Unknowns remain two — but they no longer share a parent, and the audit trail proves it.
Anchors × evidence layers
Where A5 sits.
A filled mark is primary. An open mark is secondary. A dash means the anchor does not touch that layer.
| Anchor | Intent | Exploration | Verification infra | NFR / Perf | Runtime |
|---|---|---|---|---|---|
| A1 — Test Conversion | — | — | ● | — | ○ |
| A2 — Requirements Validation | ● | — | ○ | ○ | — |
| A3 — Test-Case Generation | ○ | — | ● | ○ | — |
| A4 — Persona Bug Detection | — | ● | ○ | — | ○ |
| A5 — Spec-Driven Assurance | ● | ○ | ○ | ○ | ○ |
Open threads
What A5 asks next.
What share of evidence effort by risk tier suffices per change class?
A reusable library of constitution articles per regulation and system type.
Which drift-in-distribution signals should re-open Clarify?