Microsoft SDD
Seven phases — Constitution, Specify, Clarify, Plan, Tasks, Implement, Validate — in which intent is captured before code is generated. Spec quality equals output quality is asserted, not yet measured.
The blueprints already answer what evidence an AI-infused system needs, and whether that evidence justifies trust. What they did not yet have is an account of systems generated from a spec. Spec-Driven Assurance is that account: every Spec-Driven Development artifact carries an assurance obligation and a Confidence Engineering question.
Spine owns Constitution through Clarify, and Validate. Builders own Plan, Tasks, and Implement. NFR runs as a parallel rail from constitution to monitor.
AI Assurance answers what evidence is required. Confidence Engineering answers whether that evidence justifies trust. Spec-Driven Development supplies the missing front end: the spec as the shared source of truth for humans and AI.
Seven phases — Constitution, Specify, Clarify, Plan, Tasks, Implement, Validate — in which intent is captured before code is generated. Spec quality equals output quality is asserted, not yet measured.
AI Assurance, Confidence Engineering, and the Crosswalk. One operating rule already: every AI Assurance artifact carries a Confidence Engineering question.
Define, analyse, trace and verify, hand over, assure. It specifies and measures. It does not build, and it is not a compliance management system.
The AI Assurance lifecycle is not replaced. Scoping and test design fold into the spec and the plan — where the Intent Layer said the most cost-effective gap-closing investment was.
DeFOSPAM lens counts, IEEE 830 characteristics, and Planguage quantification put an interval on what Spec-Driven Development only asserts. Recorded at generation time and compared with output confidence later, spec quality equals output quality becomes a testable claim.
For AI-touched behaviour the criterion is a 0–2 rubric over a named, versioned set, under stated conditions, with an interval. Binary pass/fail only where Design AI Testing already says it belongs.
Generated tests are not evidence for generated code. Evidence must come from outside the generator’s distribution — Level 3 human-in-the-loop, a persona portfolio, operational telemetry — with provenance distinct and logged.
Monitoring as Continuous Evaluation already inverted the lifecycle. Converge re-reads the stream as a regression delta, and the Confidence Engineer owns the verdict.
An executable, quantified specification is the primary artifact. A5 is primary on the Intent Layer and secondary on every other layer — the first anchor to close Anchors × Evidence Layers end to end.