[leadingproblemsolver]

Target state

Ship the smallest auditable founder-advisory execution loop for the Sophie/QDB use case:

interaction -> evidence -> state change -> decision -> commitment/experiment -> follow-up -> outcome -> next-session brief

Product boundary

• no mentor AI
• no autonomous advising
• no vector DB / RAG
• no portfolio intelligence until repeated real usage
• LLM proposes structure; code controls state; evidence controls truth

Stack decision

• Next.js + TypeScript vertical under apps/qdb-advisory/
• Supabase/Postgres source of truth
• Vercel Workflows for durable orchestration, retries, sleeps, and human/event waits
• structured extraction + deterministic validation/state guards

Current artifacts

• apps/qdb-advisory/README.md — v0 execution contract, invariants, runtime flows, done-when gates
• apps/qdb-advisory/db/migrations/001_core.sql — eight-table canonical state model + idempotency + derived overdue commitments

Implementation gates

• scaffold app/runtime dependencies
• interaction ingestion API with content-hash idempotency
• structured extractor schema
• evidence support validator
• transactional object persistence
• state reducer with reconstructability test
• durable commitment lifecycle workflow
• founder update reconciliation
• on-demand next-session brief
• three-session fictional demo fixture
• hostile tests: unsupported claim, invented deadline, inferred completion, duplicate ingestion, cross-founder isolation
• precision evaluation against gold labels
• preview deployment + end-to-end receipt

Acceptance boundary

Do not merge as complete until a raw session can traverse the entire closed loop with no manual database edits and zero hallucinated critical objects.