Specification-Anchored Development — Lifecycle Architecture

From fuzzy idea to verified production through progressive ambiguity reduction — for software products

Click any phase for its role, conduct, output, and gate criteria. Press Esc to close the detail panel.

Specification-Anchored Development — Lifecycle Architecture Interactive vertical lifecycle diagram. Click any phase for detail. Specification-Anchored Development — Lifecycle Architecture From fuzzy idea to verified production through progressive ambiguity reduction — for software products DOCUMENTATION PHASES lives in ~/Documents/projects/<entity>/ (vault) Idea fuzzy · implicit assumptions · unnamed contradictions Raw sources (immutable) _ideas/ _sources/ _sessions/raw/ 21 subtypes: idea-dumps, papers, podcasts, conversations, screenshots, session proposals... fed into 0A; stress-tested in 0B Phase 0A — Externalize & Gather Eliminates: the invisible. Externalize the idea AND deposit raw sources (papers, podcasts, conversations) that ground it. ★ 0A → IDEA_MAP.md + _sources/ deposits Phase 0B — Stress-Test Eliminates: gaps in the research foundation. Finds what's missing in the idea AND in the research it rests on. If no research exists, this phase exposes that. If research is rich, this phase stress-tests it. ★ 0B → RESEARCH_SYNTHESIS.md Phase 0C — Articulate the Unconstrained Vision Eliminates: self-censorship. Includes Version Lattice + "What This Will Never Be". ★ 0C → VISION_SPEC.md Phase 1A — Scope Eliminates: "is X in this version or later?" Subtractive cut of the vision. ★ 1A → VERSION_SPEC_v1.md Phase 1B — Specify ↻ Eliminates: interpretive space the AI operates in during implementation. Spec Convergence Loop: ambiguity < 10% (critical) before ★ 1B. PRODUCT_SPEC DATA_MODEL_SPEC IMPL_SPEC BEHAVIORAL_SPEC VERIFICATION_SPEC Derivation rule: each layer derives from the one above. Parent wins on conflict. Ambiguity Catalog: unresolved ambiguities from Spec Convergence carry into Verification Spec as priority test scenarios →→→ DOCUMENTATION ↓↓↓ IMPLEMENTATION ~/Documents/projects/<entity>/ (vault) → ~/code/<repo>/ (code folder) IMPLEMENTATION PHASES lives in ~/code/<repo>/ Phase 2 — Implement Eliminates: drift from spec. Decision Boundary: CONFIRM / REDIRECT / ESCALATE on any ambiguity. ★ 2.x catalog to Phase 3 Phase 3 — Verify ↻ Eliminates: silent compliance failure. Implementation Convergence runs scenarios from Verification Spec + ambiguity catalog. Iterate until compliance ≥ 95% (generated) and 100% (cataloged) before ★ 3. Phase 4 — Spec Evolution (on verified gap) Phase 5 — Completion Full spec audit; Vision Spec compatibility check; version archival alongside code. ★ 5 Production specified · verified · versioned · eval-anchored ongoing: CI/CD + Spec Evolution (Phase 4) CI/CD · monitoring · Spec Evolution Cross-cutting disciplines · Two-layer data model: Raw (immutable) ↔ Compiled (re-buildable). Context survives AI compaction. · Scope levels: Global / Project / Local mapped to Claude Code memory. · Session-opening protocol reads SESSION_LOG + CLAUDE.md before any code. · Human ★ checkpoint at every phase gate. · Capability-skills: time-awareness, session-capture. Inspired by Andrej Karpathy's writing on LLM auto-research: without specs, the LLM has no ground truth to evaluate against — evaluation becomes circular. SAD's claim: compounding clarity (phase-by-phase ambiguity reduction) collapses compounding error.
↻ adversarial pressure-test (Convergence Loop)
Phase 4 feedback loop (on verified gap)
ambiguity catalog bridge
post-production CI/CD + evolution
★ human checkpoint