Rule: When developing INHERIT v2 + InheritKit + IAS architecture, re-derive from primary sources (product specs + library + T-files + web research) rather than amend cumulatively. BUILD-PLAN.md per repo only exists at build-start (Phase 5 closure of audit) — it is a build-driver, never a planning doc.

Why: As of 2026-04-30 Rich surfaced that incremental BUILD-PLAN amendment chains had produced “planning-doc-masquerading-as-build-driver” artefacts disconnected from the library knowledge that should ground them. Three concerns: (a) library at ~/off-github/library/ was being cited via T-file proxies; (b) BUILD-PLAN.md files accreting cumulatively rather than synthesised from coherent grounding; (c) MFI tax-service-partner-comparison feature support claimed but unverified end-to-end. Underlying failure: BUILD-PLAN was simultaneously trying to be both planning-doc AND build-driver, doing neither well.

Correction: T-files are credible primary citations (verbatim quotes + ULIDs + GRADE-CERQual). Earlier framing undersold them. The real failure was skipping BOTH T-files AND books in recent cascades, citing only prior amendments.

How to apply:

  • Every Q under new methodology grounds in library + T-file + web (cascade file lists each grounding source with verbatim quotes per feedback_actively_use_t_files_in_scorecard_authoring)
  • Strict one-at-a-time questioning (no batching tier; reverts the design-spec v1.0 §18 deviation; honours triple-elevated feedback_scorecards_one_at_a_time_optimal_sequence)
  • Detailed numerical scorecard with ≥7 sensitivity perturbations per Q (per feedback_display_numbers_in_scorecards)
  • Genuinely innovative options expected (library-driven / cross-domain / 2025-2026 SOTA / synthesis); honesty discipline if no genuine improvement found
  • Four-layer info flow per Q-answer:
    • Layer 1: per-Q cascade file at current-questions/Q-NN-topic.md (≤30s commit)
    • Layer 2: arch-state amendment (Tier-1 + Tier-2 IK + Tier-3 IAS); A-1..A-125 preserved as audit trail
    • Layer 3: per-repo DESIGN-DECISIONS.md accumulation
    • Layer 4: BUILD-PLAN.md v2.0+ at build-start only (Phase 5 closure)
  • Build-start trigger D (full pre-build-start audit closes): 76-122 Q’s; 5-8mo at Rich-pace; 7-10mo realistic. Re-open if (a) LL ARR slips past Q3 2026 / (b) competitor closes stealth window / (c) cumulative >10mo / (d) pre-acquirer signal needs earlier code.

Canonical impl: docs-strategy/docs/superpowers/specs/2026-04-30-architecture-rederivation-design.md v1.0 (commit 87660b8). Implementation plan: docs-strategy/docs/superpowers/plans/2026-04-30-architecture-rederivation-implementation.md v1.0 (commit 8acbaed).

Supersedes: docs-strategy/docs/superpowers/specs/2026-04-29-multi-phase-audit-design.md v1.0 (now v1.1 with superseded_by field). The v1.0 audit infrastructure (INDEX + 8 phase docs + Q-001 first-Q seed + STARTING-PROMPT + validate-audit-cascade.py) is partially retained / partially rewritten / partially reframed per design spec §11.

Sister disciplines (all elevated 2026-04-30 by this pivot):

  • feedback_scorecards_one_at_a_time_optimal_sequence (no batching reverts; was deviated from in design-spec v1.0 §18)
  • feedback_actively_use_t_files_in_scorecard_authoring (T-files credible primary; not proxies)
  • feedback_always_check_library_indexed_first (every Q grounds in ls ~/off-github/library/indexed/ + grep first)
  • feedback_display_numbers_in_scorecards (numerical matrix + ≥7 sensitivity perturbations per Q)
  • feedback_automatic_deep_dive_when_options_feel_imperfect (auto-trigger if margin <2pp or option-set narrow)
  • feedback_scope_vs_commercial_reality (build-start trigger D explicit re-open criteria; LL ARR cover monitored)