Where the CDI lands

What the paper argues

Delivery leads now have several AI delivery methods landing at once: SPDD, BMAD, Kiro and the broader spec-driven category. The instinct is to pick one and standardise, and that instinct is wrong. The methods are not alternatives to each other. They answer different questions and sit at different layers of the delivery system: tooling, workflow, operating model and measurement. Comparing them feature by feature produces a checklist with no architecture behind it.

The Cognitive Delivery Index sits above all of them. It does not measure whether a team has adopted SPDD or BMAD. It measures whether the governance principles those methods all aim at are being practised substantively, so the measurement question stays the same whatever workflow a team runs. Three illustrative scenarios, a single team, a programme of six teams and an organisation of forty teams, show how that works in practice.

Key points

• Each method has a centre of gravity. SPDD, BMAD and the spec-driven category sit at the workflow layer; Kiro, AWS’s spec-first IDE, sits at tooling. None was designed to address the operating model or measurement layers, and that is a gap in the discussion, not in the methods.
• The six CDI dimensions (Provenance Integrity, Supervisory Clarity, Scope Definition, Review Substance, Constraint Fidelity and Capability Calibration) are deliberately orthogonal to method choice. A team running SPDD can still score badly on Review Substance if it has automated review of generated code without a human reading it.
• In the programme scenario, a twenty-six-row comparison spreadsheet produced no answer. The working group standardised the artefact and the substance of review, and left agent orchestration, role definitions and tooling to teams.
• Mandating one method across forty teams produces compliance without coherence. Defining the operating model, mandating the substance and measuring quarterly with the CDI produces coherence in principles and diversity in workflow.
• The Level 4 to 5 gap is the hardest to close, because it is the gap between the structures existing and the structures being used substantively. The trend matters more than the absolute level.
• A starting point: score one team against the six dimensions, pick the two lowest, decide one improvement for each, leave the workflow method alone, and score again three months later.