Skip to content

Ordered implementation backlog

  1. Contracts and configuration (Pydantic and artifact slices present; stage ledger/error taxonomy pending) — Add Pydantic schemas, settings, error taxonomy, synthetic fixtures, SQLite stage ledger, and artifact-service tests. Done when invalid payloads fail deterministically and duplicate writes are no-ops.
  2. Inventory (implemented for supported formats) — Implement supported-format scanning, exclusions, iCloud placeholder detection, SHA-256 identity, and revision ledger. Done when rename, duplicate, changed, missing, and placeholder fixtures behave correctly.
  3. Native extraction (direct parsers and deterministic chunking present; Vision runtime gate pending) — Implement PDFKit/direct extraction and Swift Vision OCR adapter with page provenance, pixel caps, and resumable artifact uploads. Done when mixed PDF fixtures produce reproducible bundles.
  4. Structured extraction (local draft slice present; typed proposal pipeline pending) — Add the host Codex CLI gpt-5.6-luna distillation adapter and LM Studio OpenAI-compatible embeddings adapter, JSON schema prompts, validation, repair-once behavior, and quarantine. Done when fake and local adapters produce identical validated contracts.
  5. Retrieval and review (lexical+dense RRF, injectable reranking, and lazy Hugging Face adapters implemented; corpus indexing and held-out evaluation pending) — Add BGE embeddings, lexical+dense reciprocal-rank fusion, top-30 reranking, deterministic merge proposals, contradiction preservation, and review export. Done when held-out retrieval and no-auto-merge tests pass.
  6. Canonical publication and graph (note writer, pure projection, parameterized Neo4j rebuild adapter, and graph-rebuild CLI implemented; live rebuild and canonical-note population pending) — Add expected-hash note writer, redirects, conflict records, parameterized Neo4j upserts, relationship reconciliation, cycle detection, and rebuild command.
  7. Curriculum (pure planner slice implemented; persistence pending) — Implement prerequisite closure, topological ordering, explained 40/25/20/15 ranking, deterministic ties, and 90-minute plan allocation.
  8. Learning and Anki (mastery, CLI event, and offline export slices present; durable ledger pending) — Add event ledger, mastery dimensions, carry-forward behavior, and stable offline Anki export; AnkiConnect remains optional.
  9. Evaluation and operations (SVG reporting, Compose validation, queue worker, and documentation checks present; empirical gates/backups pending) — Add metrics, reports, chart generation, backups, recovery drills, Compose validation, documentation checks, and five-book pilot gate.

Each item is blocked by the previous item’s acceptance checks. Stop after each item with a fixture-backed review; do not scale to 100 books until the five-book gates pass.