Skip to content

Releases

PageLedger 0.9.0

Choose a tag to compare

@peterbussch peterbussch released this 15 Nov 18:22
81b6679

PageLedger 0.9.0 is the candidate for 1.0. It writes down the contract that 1.0 will keep (commands and options, exit codes, configuration keys, the adapter protocol, artifact schemas, warning codes and holds) and tests the package against it; see stability. 0.8.0 was prepared but never published, so its changes ship here too; if you come from 0.7, read both sections of the changelog.

  • Start without a config. pageledger process SOURCE --out DIR reads a PDF, page images or a text file with local tools only, shows progress, and stops cleanly on Ctrl-C with the resume command.
  • New sources. Page images and IIIF items (fetch); a library's own OCR (ALTO) can be used as a witness.
  • Reading. Dense pages read in tiles; Latin-script print (latin-print-0.1) alongside Russian; image-stage pages read concurrently in batches (from 0.8).
  • Review. An HTML review view with corrections, alongside the review sheet; reviewed jobs export only what was checked.
  • Received folders don't run code. A job, run or bundle someone sends you can't make PageLedger import code or contact a server you didn't name. Custom adapters load only from the folder given with --adapter-path, a vision server off this machine needs --trust-endpoint, and a refusal leaves the folder byte-identical. See jobs, runs and bundles you receive and SECURITY.md.
  • Verification. verify re-derives a run's normalized pages, grades and review queues, so a one- or two-file edit fails, and names the file that changed.
  • Text fidelity. Text sources keep their line endings byte for byte, right-to-left rows are no longer reversed, engine agreement keeps combining marks inside words, and numbers like 1,5 and 12 34 are recorded as conversion errors instead of being misread.
  • Single-word commands. inspect, verify, review, export, init and the rest; the earlier names still work as hidden aliases.

Before release the branch had six independent reviews (package and schemas; tests, docs and packaging; untrusted input; a newcomer installing from PyPI; text fidelity; and a review of the fixes). Each of about 130 findings was reproduced before it was fixed. Every fixture written by 0.5.2 to 0.8.0 still verifies byte for byte. Known limits are listed in capabilities and limits.

Validation: the full suite passed with and without the pdf extra (2,686 and 2,502 tests) with Ruff and mypy clean, CI passed on Python 3.10 to 3.14, wheel and source-package inventories passed, and both maintained tutorials passed against isolated installations of both packages. The wheel built byte-identically in the local build, the verify run and the publication run.

Install with pip install pageledger==0.9.0, or pip install "pageledger[pdf]==0.9.0" for PDFs; pageledger[rapidocr] adds the RapidOCR engine and pageledger[ru] the Russian lexicon. Scanned PDFs also need separately installed OCR tools. pypdf is pinned below 7, which changes extracted text.

The attached distributions and SHA256SUMS are the exact artifacts from the protected PyPI publication workflow (run 43745291786).

04e48e264a9286fe38b10fd8db543fca1ca3944a3b94a06a7c4c54ecd504716e  pageledger-0.9.0-py3-none-any.whl
7145b880ed8b27602d3cc92295690c1320d50275fc0864f8842b836b0777a22a  pageledger-0.9.0.tar.gz

PageLedger 0.7.0

Choose a tag to compare

@peterbussch peterbussch released this 02 Oct 14:55
11fde5a

PageLedger 0.7.0 adds an escalation ladder for document jobs: pages climb from local OCR to a vision model on more than warning holds, the model's reading is checked word by word against the OCR, and the differences that remain are settled against the page image, with every reading, span and decision kept as a verifiable record.

  • Climb. processing.escalate_on sends a page to the image stage when a warning holds it (as before), when its share of known words is low (rough, with the Russian lexicon from pageledger[ru]), when clean engines disagree (disagreement), or for every page (always). Each trigger is recorded on the page with its evidence, and the report shows the work done at each stage.
  • Contest. processing.contest compares a model's reading with the literal engine's and keeps each difference as a span. Look-alike letters and RapidOCR's known glyph confusions are settled by rule; a non-word those confusions don't explain stays open, because it may be the printer's misprint. Pages with open spans are held as contested.
  • Settle. pageledger adjudicate --packets writes each contested page's open spans and image, blind to which engine read what; --decisions records the answers as receipts bound to those spans. A page with no open span becomes adjudicated_text, and its edition is what transcripts and exports carry. verify-job rebuilds the spans and editions from their receipts.
  • Measured. On a hand-checked Russian article, RapidOCR differed from the checked text in 91 words, the model's reading in 17, and the adjudicated edition in 4. Across two hand-checked documents (61 pages) the rules settled 178 differences with no errors, and the spans covered 306 of the 319 known corrections.
  • Also new: a foreign_script_characters warning, reading_order: columns for the rapidocr adapter, and doctor --config reporting the keys a config names through env_key.
  • Fixed (from the unpublished 0.6.1): a scanned page with no text layer no longer waits for a person once a non-generative engine reads it without concern. New jobs record hold_policy: "0.6.1".

Validation: the full local suite passed (1,529 tests; 1,536 with the ru extra) with Ruff and mypy clean; wheel and source-package inventories passed; both maintained tutorials passed against isolated installations of both packages. New artifact fields are optional, so jobs written by 0.6.0 verify unchanged.

Install with pip install pageledger==0.7.0, or pip install "pageledger[pdf]==0.7.0" for PDFs; pageledger[rapidocr] adds the RapidOCR engine and pageledger[ru] the Russian lexicon. Scanned PDFs also require separately installed OCR tools.

The attached distributions and SHA256SUMS are the exact artifacts from the protected PyPI publication workflow (run 37019621280).

5a9d5965361af5ab1dfb2ece427e833e6889de7e384f85cab62a8e20dc179b12  pageledger-0.7.0-py3-none-any.whl
1de0500bf774db22b9a7f753e8217c5a39e24b400edceafbf8343d00132c3b0c  pageledger-0.7.0.tar.gz

PageLedger 0.6.0

Choose a tag to compare

@peterbussch peterbussch released this 01 Oct 23:37
acafda6

PageLedger 0.6.0 makes the tool usable on real collections: it adds two engines beyond Tesseract, compares engines page by page, warns about the ways text layers and vision models fail, and hardens the failure paths that real archives exposed.

  • Engines. A built-in rapidocr adapter reads scans with RapidOCR and a PP-OCRv5 recognizer you supply (the Cyrillic one keeps ѣ), and a built-in vision adapter reads pages with a vision model behind an OpenAI-compatible endpoint, keeping the exact page image it sent as evidence. Engine recipes cover Tesseract orus, Apple Vision, Surya and others.
  • Comparison. Document jobs compare the engines' readings of each page word by word. Disagreement, differing numbers, and a model reading no other engine confirms each hold the page for review.
  • Warnings. New warnings catch text layers that carry little of the page (stamp-only, digits-only, look-alike scripts, Private Use Area characters) and vision-model failures such as repetition loops, all measured on sampled real pages before their thresholds were set.
  • Hardening. Encrypted and damaged PDFs stop with a named, typed message; missing OCR tools or language packs stop before page 1; hidden files no longer enter folder runs; oversized pages render within a pixel cap; a job that reaches a limit pauses and continues with resume --raise-limit.
  • Documentation for working with a real collection: a glossary, a reference of every warning, hold and disposition, how-to guides, and engine recipes.

Validation: the full local suite passed with Ruff and mypy clean; wheel and source-package inventories passed; both maintained tutorials passed against isolated installations of both packages; the retained 0.5.1 jobs still verify. New artifact fields are optional, so artifacts written by earlier versions still verify.

Install with pip install pageledger==0.6.0, or pip install "pageledger[pdf]==0.6.0" for PDFs; pageledger[rapidocr] adds the RapidOCR engine. Scanned PDFs also require separately installed OCR tools.

The attached distributions and SHA256SUMS are the exact artifacts from the protected PyPI publication workflow (run 36914041636).

5f9a1a9a5b7d950f79f125d170fadd95aecd63d71bd9075f1ec7db1a8e92571d  pageledger-0.6.0-py3-none-any.whl
041d1fbf5434d7f99159037eb3d411d0ee91b6ff42bc496180f59f6ff2e2bf68  pageledger-0.6.0.tar.gz

PageLedger 0.5.2

Choose a tag to compare

@peterbussch peterbussch released this 12 Sep 17:14
e211d9d

PageLedger 0.5.2 fixes parent-run validation in classify --from-run.

The command now rejects incomplete page lists, duplicate or out-of-range page numbers, and page IDs that do not match their document and page. Validation finishes before classification starts or output files are written. Valid relative input paths, reversed page ordering, and missing raw evidence keep their existing behavior.

These fixes address two Copilot findings from the 0.5.1 review. The repository guide now makes GitHub-native features and Copilot guidance part of everyday development, with review checks before merge and publication. Copilot could not review the new patch because the account had reached its review quota; an independent review found no actionable issues.

Validation: 1,193 local tests passed; Ruff and mypy passed; wheel and source-package inventories passed; both maintained tutorials passed against isolated installations of both packages. Artifact schema versions and historical validation reports are unchanged.

Install with pip install pageledger==0.5.2, or pip install "pageledger[pdf]==0.5.2" for PDFs. Scanned PDFs also require separately installed OCR tools.

The attached distributions and SHA256SUMS are the exact artifacts from the protected PyPI publication workflow.

PageLedger 0.5.1

Choose a tag to compare

@peterbussch peterbussch released this 12 Sep 15:13
ebcdaea

PageLedger 0.5.1 improves document reports and closes several correctness gaps found during a wider codebase review.

  • Separate selected output from retained review concerns, while preserving verification of 0.5.0 reports.
  • Add an executable document workflow and run both maintained tutorials against installed wheel and source packages.
  • Stop adapter calls at exact budget caps, reject negative usage, and validate malformed review and route data cleanly.
  • Bind replay page selections to the original run and share package-inventory checks across release workflows.
  • Apply consistent Python formatting in a separate commit, verified to preserve every AST, and enforce it in locked CI.

Includes a 60-page source trial with per-page evidence and clear limits. Human reader trials remain pending. The live VLM tests retained provider failures and prior OCR; they did not produce a transcription.

Validation: all 1,176 tests passed with no skips. Ruff, formatting, mypy, and all 13 pull-request CI checks pass. Fresh wheel and source packages pass metadata and inventory checks, and both maintained tutorials pass from each installed package outside the checkout. The nine retained jobs, including a real 0.5.0 job, pass verification without changes to their files. No new core dependency.

Install with pip install "pageledger[pdf]==0.5.1". See the document tutorial and validation report.

PageLedger 0.5.0

Choose a tag to compare

@peterbussch peterbussch released this 11 Sep 20:23

PageLedger 0.5.0 adds document processing jobs that retain each extraction
attempt, produce a page-referenced transcript, and collect unresolved pages
for review.

  • process manages local text extraction, OCR and optional image-model stages
    under shared attempt, token and cost limits.
  • resume verifies saved page responses before reusing them. Requests with an
    unknown outcome stop for review instead of being called again automatically.
  • inspect-job, verify-job and review-job expose progress, check evidence
    and record human decisions tied to the source and selected output.
  • Image evidence records the source page, render or crop, hashes and adapter
    identity. The example transport supports configured Gemini and DeepSeek
    models through an OpenAI-compatible gateway.
  • The simplicity review removes duplicated execution and verification logic.
    It also fixes replay coverage checks, large-integer arithmetic, CSV error
    handling and strict policy thresholds.

The documentation now starts with tested examples for individual runs,
document jobs and recovery. Historical plans and local research reports have
been removed from the maintained documentation tree.

pip install pageledger==0.5.0
# Include PDF support when needed:
pip install "pageledger[pdf]==0.5.0"

Python 3.10 or later is required. Document jobs and resumable runs need a POSIX
system such as macOS or Linux. Scanned-PDF OCR also needs Poppler and Tesseract;
image-model adapters have their own optional dependencies.

Existing run artifact schemas remain at version 0.1. Core still depends only
on PyYAML. Verification checks recorded evidence; transcription accuracy and
unresolved pages still need review against the source.

The tagged source is 3a1aa9f482baff0bde2acebcad40f250aadad89d. Its development
tests use the documented dev,pdf environment. Two PDF-test dependency guards
were added to main after tagging so those tests skip when PDF support is
absent; the runtime and release packages are unchanged by that test correction.

See the documentation
and changelog.
SHA256SUMS identifies the attached distributions published to PyPI.

PageLedger 0.3.0a1 — auditable local escalation

Choose a tag to compare

@peterbussch peterbussch released this 17 Aug 02:59
fd1c1da

PageLedger 0.3.0a1 — auditable local escalation

This alpha release hardens PageLedger's evidence boundary and adds a functional
machine-level Docling adapter example without pulling Docling's ML stack into
the dependency-light core.

Highlights

  • Adds examples/docling_adapter.py for document-batched standard Docling and
    page-scoped local-VLM escalation. Pipeline, model preset, version, batch mode,
    and adapter-native warnings are retained in provenance.
  • Detects joined/run-on Latin hidden text so corrupted PDF text layers no
    longer silently receive an A signals grade.
  • Adds SHA-256 identities for raw extraction artifacts and records effective
    parser, OCR, renderer, and PageLedger runtime identities.
  • Fails closed on adapter-controlled diagnostics, malformed or escaping ledger
    paths, rerun source drift, missing raw hashes, and audit-render divergence.
  • Makes cross-run directional comparisons require matching effective extractor,
    options, PageLedger version, grading policy, evidence basis, and schema
    identity.
  • Keeps local-VLM output explicitly uncalibrated and routed to human review.

Compatibility

  • Artifact schema version remains 0.1; new fields are additive and optional
    for older artifacts.
  • Verification is intentionally stricter for raw-file integrity, lineage/source
    drift, malformed evidence, and paths that escape a run directory.
  • Core still depends only on PyYAML. Docling remains an optional machine tool;
    PDF text support remains behind the pdf extra.

Release verification

  • Reviewed source: merge commit
    fd1c1da,
    tree-identical to reviewed PR head 75ab36f.
  • PR #2 merged after three
    independent final audits found no Critical, Important, or Minor release
    blocker.
  • Main CI:
    13/13 jobs passed across Python 3.10–3.14, dependency-light and PDF-enabled
    installs, Ruff, mypy, package smoke tests, and frozen-lock recreation.
  • Verify-only release run
    passed before production.
  • Production release run
    rebuilt, tested, retained, and published the inspected artifacts through
    PyPI trusted publishing with digital attestations.
  • The exact production wheel was installed in a fresh Python 3.14 environment;
    it reported pageledger 0.3.0a1, completed a real run, and passed
    verify-run.
  • Archive inventories contain 33 wheel entries and 87 sdist entries. No PDFs,
    dogfood runs, planning state, brand source material, credentials, or
    machine-local research artifacts are included.

Production SHA-256

030f6184fd26e1938c6a0887c9fdd8ecf8da68907f72497e55be5192ede37a8d  pageledger-0.3.0a1-py3-none-any.whl
b8815ae1183b9476a8717c03207effaad0594e3af620309a991da88f170dd318  pageledger-0.3.0a1.tar.gz

Install the alpha explicitly with:

pip install "pageledger==0.3.0a1"

Full change details are in the
CHANGELOG.

PageLedger 0.2.0 — Close the route gap

Choose a tag to compare

@peterbussch peterbussch released this 17 Jul 17:02

PageLedger can now produce the route maps it already knew how to execute.

Added

  • Structural page classification: pageledger classify probes inputs or
    reuses retained evidence from a complete run and emits a route map that can
    execute unchanged through pageledger run --routes. The dependency-free
    classifier distinguishes blank, sparse, prose, table_likely, and
    unknown, records every signal and decision in an evidence sidecar, and
    supports importable classifier hooks for domain taxonomies.
  • Generation-indexed adapter escalation: run.adapter_order selects one
    adapter and option set per original/rerun generation. Manifests record the
    planned chain and current step; exhausted chains leave pending pages in the
    human review queue.
  • Budget alerts and grouped cost evidence: capless warn_pages,
    warn_tokens, and warn_usd thresholds record one first-crossing alert per
    unit. cost.json can now group extracted-page usage and resolved cost by
    adapter and routed page type.

Compatibility

  • The package version is 0.2.0, while every artifact and config contract keeps
    schema_version: "0.1". Manifest escalation fields and cost alerts/rollups
    are additive and optional; existing 0.1 artifacts remain valid.
  • Core still depends only on PyYAML. The classifier adds no bundled model,
    computer-vision runtime, provider SDK, pricing catalog, or domain taxonomy.

Install

pip install -U pageledger

See the classifier guide for the classify-review-run workflow.

PageLedger 0.1.7 — Unicode-safe quality evidence for multilingual archives

Choose a tag to compare

@peterbussch peterbussch released this 13 Jul 14:48

Multilingual quality evidence is now Unicode-script safe.

Fixed

  • Quality metrics tokenize on Unicode letters together with their combining
    marks, eliminating false fragmented-text and symbol-noise warnings on clean
    Devanagari, Bengali, Gujarati, Gurmukhi, Tamil, Telugu, Kannada, and
    Malayalam prose.
  • Regression coverage now spans those Indic scripts plus Arabic, Kazakh
    Cyrillic, Latin, and Russian — while the existing OCR-fragment fixtures
    still warn as they should.
  • Per-page word_count now uses the same Unicode tokenizer as the lexical
    quality metrics instead of disagreeing with them on combining-mark scripts.

Packaging

  • All seven JSON Schema contracts ship inside built wheels under
    share/pageledger/schemas; CI and the publish workflow verify wheel
    contents.

Compatibility

  • Artifact schema version remains 0.1; no field was added, removed,
    renamed, or retyped. Corrected word_count, suspicious_symbol_count,
    and lexical metrics can differ from 0.1.6 for scripts that use combining
    marks or non-ASCII punctuation.

Install

pip install -U pageledger

Testing a collection we haven't seen? Open a corpus report.

PageLedger 0.1.6

Choose a tag to compare

@peterbussch peterbussch released this 11 Jul 01:27

PageLedger 0.1.6 adds executable reviewed route maps, resilient long-run failure recovery, corrected policy and cost compatibility behavior, and stronger release verification. See CHANGELOG.md for the complete release notes.