CHANGELOG.md

Changelog

Unreleased

• Streamlined the public documentation tree while preserving historical release records in GitHub Releases, Git Tags and this changelog.
• Added a private security-reporting policy and pull-request template, refreshed public issue forms for v1.0.0-rc1, and clarified the active RC-to-stable release gate without changing runtime behavior.

v1.0.0-rc1 — 2026-08-11 — Public Release Candidate

• Recorded the three completed frozen-checklist evaluations and completed a separate unseen split-caption/figure-plate PDF after a generic crop fix for v1.0.0-rc1 preparation.
• Prevented duplicate provenance labels when structured content already begins with its declared [分析], [假设] or [用户] marker.
• Distinguished an original-page-corrected range symbol from an unchanged ± that was explicitly verified on a PyMuPDF page render.
• Added a provenance-preserving crop route for captions separated from figure plates on different physical PDF pages.
• Added a direct local-PDF to standalone-Markdown run plan that does not require Zotero or Obsidian configuration.
• Defined created as the first successful note-generation date, independent of Zotero-added, PDF-import, acquisition, and run-start dates.
• Fixed Markdown as the sole formal note output and removed deferred PDF-note and video-module claims from the v1 product scope.
• Introduced LitAnchor Notes · 文锚笔记 as the public product name while retaining the compatible litanchor-paper-reading Skill slug and repository.
• Recorded user acceptance of the final unseen note and both split-page figure crops before freezing the RC.

v0.6.0-beta.3 — 2026-08-10 — Note Experience & Paper Template v1.0 Freeze

• Froze the reader-facing content contract as Paper Template v1.0, split from the private slot-driven Runtime renderer and deterministic validation layer.
• Defined mode scope: skim renders Section 1, deep Sections 1–6, and internalize Sections 1–8; excluded sections are omitted.
• Froze the 16-field note properties, source-only keyword handling, date format, and one-mode-tag contract while keeping complete source identity in sidecars.
• Replaced visible Evidence IDs, repeated excerpts, and the evidence index with compact linked p.x locators; audit evidence remains private and complete.
• Added direct key-visual embeds with source subsection, selection reason, and a two-to-four-sentence interpretation.
• Added argument-aware paper lenses, structured experiment evidence chains, explicit conclusion boundaries, and visible provenance labels.
• Added source-grounded internalize research-idea gates: falsifiable hypothesis, delta, validation design, failure modes, and novelty status.
• Added page-grounded, structure-grounded, and source-limited delivery semantics; formal deep/internalize rendering requires the page-grounded path.
• Removed the duplicate note H1 and added readable, stable filenames for long or filesystem-unsafe paper titles.
• Required versioned bundled tools during a paper run; failures are preserved instead of being hidden by inline replacement scripts or mid-run Skill edits.
• Kept skim visual inventories in private sidecars without requiring or copying image embeds that are intentionally excluded from the Section 1-only note.
• Removed internal HTML comments and Section 6.5 from reader-note source, while keeping validation and evidence details in private sidecars.
• Rendered supporting reproducibility records without leaking the internal core-experiment completion placeholder.
• Added adjacent-caption crop boundaries so vertically joined figures are separated before PyMuPDF renders the selected visual.
• Made text-heavy Markdown tables explicitly left-aligned without requiring Vault CSS.

v0.6.0-beta.2 — 2026-08-01 — Public Beta Stabilization

• Unified manual and autonomous PDF preparation on PyMuPDF, removed the redundant pypdf runtime dependency and made the documented physical-page authority match the implementation.
• Added a shared coordinate-aware PyMuPDF reading-order pass that preserves spanning headings and reads each detected column from left to right.
• Changed core, optional MinerU and development requirements to minimum-only constraints while preserving their lightweight installation boundaries.
• Rebuilt the bilingual landing-page pipeline and detailed workflow as centered, staged Mermaid diagrams so Section Synthesis remains readable on GitHub and in Obsidian.
• Split public failure reports into installation/lifecycle, runtime/PDF and note-quality issue forms with explicit privacy and anti-leak confirmations.
• Added an allowlist-built doctor --support-bundle ZIP containing safe versions, check statuses, stable error codes, MinerU routing and receipt metadata without paper data, identifiers, credentials or local paths.
• Added multilingual semantic intent-regression cases across all lifecycle operations, five one-paper selector forms and ambiguity handling without introducing phrase matching into the runtime Skill.
• Added interrupted-activation repair, configuration preservation, Vault non-interference and real Windows lifecycle smoke coverage.
• Added a release-time Skill compaction gate for entry length, duplicate long rules and missing referenced resources.
• Reworked the bilingual repository landing pages around the two-intent public experience, reliability contract, current boundaries and public-beta proof.
• Added a current documentation index, stable-release roadmap, contribution guide and privacy-aware GitHub issue forms.
• Removed obsolete v0.4.1 previews, paper figure crops, failed sparse outputs, repair reports, the superseded early project specification and raw personal reading/template source notes from the current branch. Git tags and releases continue to preserve version history.
• Kept the v0.6 evaluation manifests, frozen holdouts, pathological fixtures, rubrics, metrics and validation reports unchanged.
• Clarified that the distilled reading method is an optional project-authored heuristic for internalize, not a factual source or reliability rule.

v0.6.0-beta.1 — 2026-07-29 — Zero-Config Public Beta

• Synchronized the current product, integration, workflow, schema, evaluation and release-checklist documents with the published v0.5.0 state.
• Documented the supported Windows/Python/Zotero/Obsidian baseline and the minimum-version, capability-probed compatibility policy.
• Added the v0.6 Agent-assisted installation and public-beta hardening roadmap.
• Removed unused integration comparisons from the active Skill reference and current user-facing boundaries while retaining historical specifications as clearly marked records.
• Added a machine-readable installation manifest, receipt-tracked user-local installer, idempotent install/upgrade/repair, archived rollback and confirmation-gated uninstall.
• Added Obsidian Vault discovery by name, contained first-run setup, persistent MinerU consent, machine-readable doctor checks and an Agent-facing run plan.
• Added deterministic Windows release ZIP/checksum generation and Windows Python 3.10–3.14 GitHub Actions coverage.
• Made Agent entry intent-based instead of requiring two literal example sentences, and added provisional support for future Python/Zotero versions.
• Kept the public dependency split lightweight: two core PDF packages, one optional MinerU requirement and PyYAML only for development validation.
• Passed three release-eligible frozen holdouts under the pre-published v0.6 rubric after moving two failure-driven cases to development and disqualifying one previously exposed replacement.
• Completed whole-document and page-preserving MinerU routes on the holdouts: 236 of 265 blocks aligned to PyMuPDF pages, two section candidates and twelve figure candidates were added, and zero authoritative EvidenceUnits came from MinerU.
• Passed six synthetic pathological PDF/failure cases, including two-column order, scan fallback, mixed blank pages, long documents, Methods-after-References and MinerU timeout fallback.
• Expanded the anti-leak audit to include titles and DOIs from every tracked evaluation manifest rather than only the original v0.5 corpus.
• Made GitHub Actions fetch complete Git history so frozen-evaluation ancestry checks remain reproducible on pull-request runners.

v0.5.0 — 2026-07-26 — Autonomous Deep Reading

• Promoted the autonomous deep-reading pipeline after the release candidate, six-paper evaluation and three additional cross-domain generalization cases.
• Passed a final fully unseen, source-closed smoke test under the frozen workflow and rubric with no severe fidelity failure.
• Kept the release paper-independent: no test title, answer, reference note or paper-specific branch was added to the distributable Skill.
• Preserved PyMuPDF as the physical-page and evidence authority while keeping MinerU Flash as a consent-controlled, non-authoritative structure enhancer.
• Kept visual acceptance, no-overwrite Obsidian export, local-only consent and runtime privacy boundaries unchanged.
• Stable release gate: 116 repository tests, Skill validation, Python compilation and anti-leak/privacy checks passed locally.

v0.5.0-rc1 — 2026-07-26 — Autonomous Deep Reading Pre-release

• Froze a three-paper blind corpus using only Zotero metadata, original PDFs and the canonical Final template.
• Added a PyMuPDF-authoritative autonomous work-packet command with automatic paper-type classification, section discovery and six explicit reading passes.
• Added blind-run origin controls: only auto_extracted EvidenceUnits and auto_synthesized ClaimRecords can enter an autonomous candidate.
• Added separate fidelity and recall review contracts. Pending or incomplete reviews cannot represent a completed deep-reading run.
• Added token-free MinerU Flash routing for whole eligible papers and page-preserving subsets of long PDFs; all MinerU content remains non-authoritative until aligned to an original PyMuPDF physical page.
• Added three persistent local MinerU consent modes (always_for_eligible_files, ask_each_time, never) and automatic structure fusion for eligible files; the selected mode is never committed.
• Added page-verified autonomous semantic materialization, complete SectionSynthesis, independent fidelity/recall review, all-figure selection, crop validation and Final-template candidate generation.
• Added explicit page-review receipts and preserved already-reviewed pages when a later MinerU fusion leaves their physical-page classification unchanged.
• Fixed Nature-style Methods-after-References layouts, visual appendices after references, Fig. N | captions, legitimate mm units and accepted-note export promotion.
• Completed the official six-paper autonomous evaluation with 182 EvidenceUnits, 155 ClaimRecords, 15 selected visuals, zero blockers and all deterministic quality metrics at 1.0.
• Rebuilt ResNet, LOVECLIM and climate U-Net as extended non-autonomous regressions; the stricter summary gate caught and blocked a shallow LOVECLIM overview until its evidence-backed summary was corrected.
• Added cross-paper metrics, failure taxonomy, anti-leak audit, release checklist and a non-promoted EvolutionCandidate.
• Added paper-independent checks for range-symbol corruption, metadata/type consistency, generic claim headings and duplicated section content.
• Added primary plus secondary paper profiles, a distinct parameter claim type, original-page scientific-symbol receipts, multi-evidence combined claims and idempotent acceptance of reviewed warning-state notes.
• Compacted SKILL.md to the permanent workflow contract; detailed rules, deterministic behavior and templates remain in references/, scripts/ and assets/.
• All six official notes passed user review for key visuals, layout and Zotero links. Three additional unseen cross-domain papers also passed structured regeneration and user review without adding paper-specific rules.
• Full repository regression: 116 tests passed; Skill validation, compilation, privacy and anti-leak audits passed.

v0.4.1 — 2026-07-25 — Deep Reading Pipeline Pre-release

• Reclassified the previous ResNet, LOVECLIM and U-Net outputs as failed deep-reading examples rather than accepted blind notes.
• Added the canonical slot-driven Paper Template - Final asset for all deep and internalize notes.
• Expanded ClaimRecord types and added structured titles, explanatory details, conditions, section IDs and importance.
• Added deterministic deep-reading gates for required content groups, content recall, section depth and Final-template conformance.
• Kept skim output separate from the Final deep-reading template.
• Added regression tests proving that a sparse one-claim ledger is blocked and a rich ledger renders the complete Final template.
• Reprepared and rebuilt ResNet, LOVECLIM and climate U-Net as real-paper forward-validation cases; all three now pass Final-template, multi-page traceability, content-recall, section-depth, quote-completeness, numeric-rendering, cross-section consistency and paper-type gates.
• Fixed duplicated numeric suffixes, rejected incomplete evidence quotations, mapped data/metrics/experiments into the correct Final-template sections and required an all-figure inventory with both method and result coverage where available.
• Added result visuals for the three repaired examples: ResNet Figure 6, LOVECLIM Figure 21 and U-Net Figure 7.
• Published the release as a GitHub Pre-release because the local runtime still requires the specialised reading passes to populate the ledgers.
• Automatic multi-pass full-text deep reading, automatic PyMuPDF–MinerU fusion and independent semantic review are explicitly deferred to v0.5.