Sections 7 and 8
Adoption, right-sized
SDD’s playbook, the blueprints’ order, and only the pieces of this extension that belong at each step.
| Step | SDD playbook | Blueprint order | What to adopt from this extension |
|---|---|---|---|
| 1 | Pilot — one feature with visible alignment pain | Start with the most-upstream anchor (A2) — lowest cost, highest leverage | The spec template and the Spec Confidence block; Clarify as a named gate |
| 2 | Formalize — a lightweight spec: scenarios, constraints, acceptance criteria | Add verification infrastructure (A1, A3) | The assurance constitution; distributional criteria; the plan’s assurance gate; [E] tasks |
| 3 | Iterate — AI generates artifacts from that context | Add exploration (A4) | Provenance logging; Level 3 HITL and persona exploration scheduled against [E] tasks |
| 4 | Refine and scale | Promote to runtime; cross-cut with NFR | The validate-and-converge record; promoted monitors; converge after each increment; the spine where many builders share one definition |
Right-size. A patch carries the regression set and the promoted monitors. A light change carries its coverage areas’ [E] tasks. A full change carries the whole budget. Probabilistic acceptance, provenance, and a named owner of the verdict apply to every class.
Measures
Each held to the recursive standard.
These are not activity metrics. Each one, when reported, still has to answer how confident we are that the number means what we think it means.
| Measure | What it tells you | Why it is not an activity metric |
|---|---|---|
| Spec confidence index | Lens counts and IEEE 830 scores with an interval, per spec version | A distribution, not template completion |
| Open clarifications | Count, age, and blocking status of markers, by lens and coverage area | Measures what is not yet known |
| Evidence budget by risk tier | Share of [E] effort on the risks that would hurt most | Effort follows risk, not feature count |
| Output confidence | Rubric distributions with intervals per coverage area, under stated conditions | “At least 90% score 2 across N runs”, not “test passes” |
| Provenance independence | Share of AI-touched flows with at least one evidence source distinct from the generator | The Two Unknowns, counted |
| Regression delta | +1 / 0 / −1 per prompt per converge; drift in the distribution | A distribution of change, not a pass |
| Promoted-monitor coverage | Share of high-risk scenarios with a live monitor and an alert owner | Evidence that is current |
| Handover rework | Builder clarifications raised after handover | The direct proxy for translation losses 2 and 3 |
| Spec confidence vs output confidence | The two indices side by side over time | Makes spec quality = output quality testable |