SDD ModeAIAD Framework

Interactive overview

The SDD Mode Big Picture

The whole operating mode on a single diagram: the seven-step cycle, the Tier 1 governance band above it, the artifacts that feed it, the responsibilities that carry it and the metrics that measure it. Click any element to open the corresponding article.

Big Picture of the SDD Mode cycle — from intent to chain of evidence Interactive diagram: a Tier 1 governance band with veto power sits above the seven-step SDD cycle (Intent Statement, Research, SPEC, Execution Gate, Agent Execution, Validation, Drift Lock), followed by deployment and consolidation which loop back to the intent. Above the Gate and execution, an “Agent autonomy” marker recalls that the Gate also checks the execution perimeter and the stop threshold whenever an action is irreversible or an agent holds credentials. Validation relies on designed verification; the Drift Lock is the last link of the chain of evidence that conditions deployment. Below: the artifact hierarchy, the five responsibilities, the rituals and metrics bands, and Human Authorship expressed as an authority continuum. A full text alternative follows the diagram. TIER 1 GOVERNANCE — veto power · fail-closed: UNKNOWN = VETO AIAD-AI-ACT — EU AI Act AIAD-AI-ACT Regulation (EU) 2024/1689 AIAD-RGPD — Privacy by Design AIAD-RGPD GDPR (EU) 2016/679 AIAD-RGAA — Accessibility AIAD-RGAA RGAA 4.1 / WCAG 2.1 AIAD-RGESN — Eco-design AIAD-RGESN RGESN v2 eco-design THE SDD CYCLE — no step skipped Agent autonomy — the Gate checks the execution perimeter and the stop threshold for irreversible actions and credentials Agent autonomy: perimeter · stop threshold 1. Intent Statement — capture the human WHY STEP 1 Intent Statement /sdd intent 2. Research — intent viability, GO / NO-GO verdict STEP 2 Research /sdd research · GO/NO-GO 3. SPEC — living specification, prose or EARS STEP 3 SPEC /sdd spec · /sdd split 4. Execution Gate — SQS ≥ 4/5 required; for irreversible actions and credentials, also the autonomy ceiling, execution perimeter and stop threshold STEP 4 · GATE SQS ≥ 4/5? /sdd gate < 4/5: back to revision 5. Agent Execution — optimised context, plan before code, agent-initiated stop points STEP 5 Agent Execution /sdd exec · /sdd resume 6. Validation — four axes (technical, functional, business, governance); verification is designed, not bought STEP 6 Validation /sdd validate · Tier 1 veto designed verification 7. Drift Lock — code and SPEC in the same PR, last link of the chain of evidence STEP 7 Drift Lock /sdd drift-check · /sdd trace last link of the evidence 8. Deployment — staging then production, monitoring; automated only when the chain of evidence is complete STEP 8 Deployment staging → prod · rollback on a complete chain of evidence 9. Consolidation — retro, Intent Workshop, status STEP 9 Consolidation /aiad retro · /aiad intention next iteration JNSP — uncertainty policy: say "I don't know" rather than invent JNSP — uncertainty is a valid signal, not a failure Machine-verifiable traceability: @intent, @spec, @verified-by, @governance Traceability @intent · @spec · @verified-by · @governance Context Engineering Budget — at most 60-70% of the window Context Budget 60-70% THE ARTIFACTS — documentary hierarchy PRD — product vision, injected at framing time only PRD framing only ARCHITECTURE — technical standards, permanent condensed ARCHITECTURE permanent condensed AGENT-GUIDE — permanent context, Lessons Learned, Human Learnings AGENT-GUIDE permanent context Intents — the WHY archived in .aiad/intents/ Intents .aiad/intents/ SPECs — the HOW, a single active one per session SPECs one per task THE FIVE RESPONSIBILITIES Product Manager — owner of value Product Manager value Product Engineer — guardian of intent Product Engineer guardian of intent Agents Engineer — AI ecosystem and Harness Engineering Agents Engineer the AI ecosystem QA Engineer — quality built in from the start QA Engineer quality Tech Lead — long-term technical coherence Tech Lead technical coherence Rituals and synchronisations — standup, demo, tech review, retro, Intent Workshop RITUALS & SYNCHRONISATIONS /aiad standup · demo · tech-review · retro · intention Metrics — SQS, DORA, Flow, dashboards METRICS SQS · DORA · Flow · /aiad dashboard Human Authorship — authorship of intent cannot be delegated; authority continuum: each level is a distinct human act, answering is not consenting HUMAN AUTHORSHIP value 7 · authority continuum

Tier 1 governance (veto) Cycle steps Gates & Human Authorship Artifacts & Drift Lock Responsibilities Rituals & metrics

The cycle in text

Full alternative to the diagram above:

  1. Intent Statement (/sdd intent) — an identifiable human captures the WHY in five fields. Without a human author the session is invalidated (Human Authorship). The value unfolds as an authority continuum: each step up — clarification, SPEC validation, write authorisation, deployment authorisation — is a distinct human act, and answering the agent is not consenting.
  2. Research Phase (/sdd research) — mandatory code-grounded discovery, graduated verdict GO / CONDITIONAL GO / DEFER / NO-GO decided by the human.
  3. SPEC (/sdd spec, /sdd split) — the living specification: scope, interface, behaviour, edge cases, tests, drift criterion. Prose or EARS format.
  4. Execution Gate (/sdd gate) — the SQS scores 5 criteria plus the Stranger Test. Below 4/5, the gate closes and the SPEC goes back to revision. As soon as an agent action is irreversible or an agent holds credentials, the Gate also checks the autonomy ceiling, prior human authorisation, the execution perimeter and the stop threshold (see Agent autonomy).
  5. Agent Execution (/sdd exec, /sdd resume) — optimised context within the Context Engineering Budget, implementation plan required before code. The agent stops on its own at the stop points set by the SPEC.
  6. Validation (/sdd validate) — four axes: technical, functional, business, governance. The 4 Tier 1 agents hold a fail-closed veto. Verification is designed, not bought: what is checked, when and by whom matters first.
  7. Drift Lock (/sdd drift-check, /sdd trace) — code and SPEC synchronised in the same PR, verified by the traceability matrix. It is the last link of the chain of evidence.
  8. Deployment — staging then production, monitoring, rollback if needed (Integrate loop). Automated deployment is admissible only when the chain of evidence is complete: Intent Statement, SPEC validated at the Gate, traced verifications, Drift Lock.
  9. Consolidation (/aiad retro, /aiad intention, /aiad status) — Lessons Learned, Human Learnings, metrics. Then a new iteration from step 1 (rituals).

In permanent support: the artifacts (PRD, ARCHITECTURE, AGENT-GUIDE, Intents, SPECs), the five responsibilities, the JNSP uncertainty policy, agent autonomy (authority continuum, stop points, execution perimeter, stop threshold) and the SQS, DORA and Flow metrics.