Skip to content

LitAnchor v0.5.0 — Autonomous Deep Reading

Latest

Choose a tag to compare

@xieyf1024 xieyf1024 released this 26 Jul 12:26
· 24 commits to main since this release
94d7974

LitAnchor v0.5.0 — Autonomous Deep Reading

LitAnchor v0.5.0 promotes the release-candidate workflow to the first stable
autonomous deep-reading release.

What is stable

  • Zotero Local API lookup with unique-item and unique-PDF resolution.
  • PyMuPDF-authoritative physical pages, quotations, coordinates, links and
    original-PDF figure crops.
  • Six-pass, paper-type-aware full-text reading before Final-template rendering.
  • Evidence-first semantic materialization: EvidenceUnits precede ClaimRecords.
  • Full SectionSynthesis plus independent fidelity and recall reviews.
  • Complete figure inventory, 1–3 selected method/result visuals and crop
    provenance gates.
  • Persistent local MinerU consent modes and automatic Flash routing for
    eligible files; MinerU content remains non-authoritative until aligned to the
    original PDF.
  • Hash-verified, path-contained, no-overwrite Obsidian candidate/final export.

Release evidence

  • Six official autonomous papers: 81 physical pages, 182 EvidenceUnits,
    155 ClaimRecords and 15 selected visuals.
  • Three additional unseen cross-domain papers passed structured regeneration
    and user review without paper-specific Skill rules.
  • One separate fully unseen, source-closed final smoke test passed the frozen
    rubric at 98/100 with zero predefined severe-failure signals.
  • 116 repository tests, Skill validation, compilation and privacy/anti-leak
    checks passed locally.

The final smoke paper, its PDF, generated note and runtime ledgers are not
committed. Stable code promotion does not convert pending user visual review of
that private candidate note into an accepted review.

Boundaries

v0.5.0 does not add Zotero MCP, Zotero write-back, bidirectional synchronization,
paid MinerU APIs, local full MinerU, multi-paper reviews or unattended Skill
publication. External MinerU upload still follows the user’s local consent
mode, and PyMuPDF remains the evidence authority.