Making AI-assisted delivery visible from initiative intent to production evidence.

The delivery model

The site was treated as one initiative. Product framing established the brief, architecture, shared decisions, and initial feature sequence. Each component specification then operated as a mini-sprint with a bounded outcome, stable acceptance criteria, implementation, and verification evidence.

The original brief identified seven increments. Delivery ultimately produced twelve implemented specifications and one governed design spike. Added work included lead persistence, assistant hardening, approved-claims integration, security controls, and accessibility remediation.

Making progress visible

Traditional status reporting would have reduced the initiative to a single completion percentage. The reconstructed dashboard instead separates portfolio outcomes, program flow, feature allocation, and squad execution.

• Initiative and feature progress by specification
• Spec-by-spec completion timing and 78 acceptance criteria
• Daily commit, specification, and deployment throughput
• Modeled development-token allocation by feature
• DORA-style deployment and recovery evidence

Production evidence

Vercel recorded 39 production-target deployment attempts from August 23 through August 26, 2026. Thirty-six reached READY and three ended in ERROR. The median Git-push-to-production-READY time was 23.3 seconds across the 35 successful deployments with complete timestamps.

The three failed attempts formed two observable build-failure clusters. The average recovery proxy was 6 minutes 27 seconds. Because the failures were caught in the deployment pipeline, they are not represented as confirmed customer-facing incidents.

What the dashboard exposed

The initiative reached its intended outcome and the custom domain went live, but only five of twelve component-spec status fields said Implemented. Delivery moved faster than specification-status maintenance.

That 42% spec-truth-health signal is exactly the kind of governance weakness a useful SDD dashboard should expose. The point is not to celebrate document production; it is to show whether intent, implementation, verification, and operating evidence remain aligned.

Development economics

The initiative's development effort is modeled at approximately 650,000 AI-development tokens, with a planning range of 390,000 to 910,000. The allocation considers specification work, implementation concentration, touched components, tests, review, correction, and redeployment.

The figure is an analytical estimate rather than provider billing telemetry. The AI-discovery feature accounts for the largest modeled share because it accumulated the most integration, hardening, privacy, fallback, testing, and redeployment work.