pageledger 0.1.1

Page-denominated document extraction run ledger

Record, route, and review document extraction — one page at a time.

PageLedger is for researchers who extract text from archives at scale — digital humanities labs, libraries, historians, and social scientists running OCR/VLM over thousands of scanned pages who must be able to answer, months later: which pages went through which engine and prompt, what it cost, which pages are suspect, and what happened when they were re-extracted.

It is a page-denominated run ledger — not an OCR model, not a PDF-to-Markdown converter. You bring the extraction engine (Tesseract, Docling, Marker, a cloud VLM); PageLedger routes pages to it, enforces page/token/dollar budgets, and writes the evidence as plain files: provenance, quality diagnostics, cost with its basis, audit queues, and executable rerun plans. No service, no database — a run directory you can grep, cite, and reproduce.

PageLedger grew out of soviet-corpus, a Soviet census digitization and corpus-linguistics project where those questions had to be answerable for source citation and methodological review. The run-ledger layer was extracted from that working pipeline and rebuilt as a standalone, domain-neutral tool.

The current alpha covers the run-ledger core: built-in text, pdf_text (born-digital), and pdf_ocr (scanned, via local Tesseract) adapters, custom adapters via import strings, dry runs, budgets, quality signals, page-scoped reruns (pageledger rerun), and cross-run diffs (pageledger compare-runs). It does not call network VLMs out of the box, classify pages, align records to schemas, or compute audit grades — see Current Runtime Capabilities.

Install

pip install pageledger

Optional born-digital PDF text extraction:

pip install "pageledger[pdf]"

Quickstart

Create a text fixture and a starter config:

printf 'first page\fsecond page\n' > sample.txt
pageledger init-config --out pageledger.yml

Inspect the run without extraction:

pageledger run sample.txt --config pageledger.yml --out runs/dry --dry-run

Run the built-in text adapter, then summarize the run directory:

pageledger run sample.txt --config pageledger.yml --out runs/text --json
pageledger inspect-run runs/text

For a born-digital PDF, install the optional PDF extra and use pdf_text:

pip install "pageledger[pdf]"
cat > pageledger-pdf.yml <<'YAML'
schema_version: "0.1"
taxonomy:
  page_types:
    prose:
      default_action: transcribe_text
run:
  adapter: pdf_text
YAML
pageledger run document.pdf --config pageledger-pdf.yml --out runs/pdf-dry --dry-run
pageledger run document.pdf --config pageledger-pdf.yml --out runs/pdf-text --json

--out must point to a new empty directory. pdf_text extracts an existing text layer and does not OCR scanned images.

For a scanned PDF with no text layer, use pdf_ocr. It shells out to pdftoppm and tesseract, so install poppler and Tesseract first (pageledger doctor checks for both):

pageledger init-config --adapter pdf_ocr --out pageledger-ocr.yml
pageledger run scan.pdf --config pageledger-ocr.yml --out runs/scan-ocr

Set DPI and language under run.adapter_options:

run:
  adapter: pdf_ocr
  adapter_options:
    dpi: 400
    lang: eng+deu

The run directory contains manifest.json, route-map.yml, audit.json, audit.md, provenance.jsonl, quality.jsonl, cost.json, run.log, rerun-manifest.yml, and per-page raw output under raw/.

For a worked example on a real declassified scan, see docs/examples/jfk-scanned-archive.md. For choosing between local, cloud, and hybrid extraction, see docs/ocr-options.md. Custom adapter examples live under examples/; load them with --adapter-path:

pageledger run scan.pdf --config custom.yml --out runs/custom \
  --adapter-path examples

How a Run Works

flowchart TD
    A["inputs (text / PDF)"] --> B["paginate<br/>(form-feed or PDF pages)"]
    B --> C["route pages<br/>route-map.yml"]
    C --> D{"budget preflight<br/>max_pages"}
    D -- over cap --> X["refuse: nothing written"]
    D -- ok --> E["extract page via adapter<br/>(retry + backoff)"]
    E --> F["quality signals<br/>quality.jsonl"]
    E --> G["provenance.jsonl<br/>+ cost.json (cost_basis)"]
    E --> H{"budget mid-run<br/>pages / tokens / USD"}
    H -- over cap --> Y["halt: manifest status=failed,<br/>partial artifacts consistent"]
    H -- ok --> E
    F -- warnings --> I["review queue<br/>audit.json / audit.md"]
    I --> J["rerun-manifest.yml"]
    J -- "pageledger rerun<br/>(stronger adapter)" --> E
    G --> K["pageledger compare-runs<br/>(parent vs rerun)"]

Every box on the right is a plain file in the run directory; nothing needs a service or database.

Current Runtime Capabilities

Built-in and tested:

• pageledger run for text fixtures (form-feed pagination), born-digital PDF text layers (via pageledger[pdf]), and scanned PDFs (pdf_ocr, using locally installed pdftoppm + tesseract).
• Adapter options (run.adapter_options) passed to built-in and custom adapter constructors, and --adapter-path for loading custom adapter modules without touching PYTHONPATH.
• Dry-run mode that generates full run artifacts without calling extractors.
• Per-page provenance (provenance.jsonl), quality diagnostics (quality.jsonl), cost rollups (cost.json), structured run logs (run.log), and audit/review queues (audit.json, audit.md).
• Page-denominated budget enforcement (pages, tokens, dollars) with preflight and per-page caps.
• Retry with configurable max_retries and optional exponential backoff (retry.backoff: exponential).
• pageledger rerun — re-extracts exactly the pages listed in a previous run's rerun manifest (typically with a stronger adapter), preserving page ids, recording parent lineage, enforcing max_rerun_depth, and warning if a source file changed since the parent run.
• pageledger compare-runs — page-by-page diff of two runs: character/word deltas, warnings resolved or introduced, adapters, and cost.
• Cost provenance: cost.json records cost_basis (adapter_reported, configured_rate, mixed, or none) so derived accounting rates are never mistaken for provider-billed spend, plus runner-measured extraction_seconds per page and in total.
• pageledger doctor diagnostics for optional PDF/OCR/cloud tooling.
• Configurable custom adapters via module.path:object import strings.

Adapter-supported but user-supplied:

• OCRmyPDF preprocessing, or any external engine wrapped as a custom adapter.
• Cloud OCR/VLM adapters (user provides API keys, adapter code, and pricing).
• Local document-conversion engines (Docling, Marker, Surya) through custom adapters.
• PDF page counting for custom adapters that expose page_count(source).

Documented future work (not yet implemented):

• Automatic page classification (the alpha routes every page to review in dry-run mode; no classifier ships).
• Schema alignment execution (the schema config section is parsed but the aligner does not yet produce normalized records).
• Audit grading (review queues are populated but no grades are computed).
• Multi-adapter routing chains, rerun_if/quarantine_if policy enforcement, and staged CLI commands (classify, extract, align, audit).

Known Limits

• pdf_text reads existing text layers. It does not perform OCR. For scanned PDFs use pdf_ocr (local Tesseract) or wrap a stronger engine (OCRmyPDF, Docling, Marker, Surya, cloud OCR) as a custom adapter.
• pdf_ocr needs poppler and Tesseract installed. PageLedger never installs OCR engines; it shells out to pdftoppm and tesseract and fails with an install hint (pageledger doctor) when they are missing. OCR quality is Tesseract's, at the DPI and language you configure.
• Quality signals are diagnostic, not calibrated. quality.jsonl records per-page evidence: character/word counts, replacement characters, control characters, suspicious symbol density, lexical shape (token counts, mean token length), and embedded-text deltas. A seven-item warning taxonomy (empty_text, short_text, replacement_characters, control_characters, suspicious_symbol_density, fragmented_text, suspicious_embedded_text_delta) flags pages for review. These signals are NOT OCR accuracy scores — they are evidence a human should weigh. In particular, no heuristic here detects word-level misrecognition ("matericl" for "material"); that takes a dictionary, a model, or a human, and PageLedger ships none of them.
• Quality-warning pages appear in the audit review queue. For execution runs, every page with one or more quality warnings is added to audit.json → review_queue with reason "quality_warning". Dry-run review entries use route-based reasons ("no_classifier_available").
• No automatic page routing. Every page routes to the configured default_action (or review in dry-run mode). Projects that need page-type-aware routing must classify pages outside PageLedger or wait for a future classifier.
• Schema alignment does not execute. The schema config section is validated and preserved in the config snapshot, but the runner does not yet align extractor output to declared columns, types, or checks.
• Reruns re-extract listed pages; they do not merge results. rerun-manifest.yml lists pages from audit.json → review_queue; pageledger rerun re-extracts exactly those pages into a new run directory with parent lineage. rerun_status is "executable", "empty_queue", or "no_further_generations" (depth cap reached). Combining parent and rerun outputs into one corpus remains the project's decision — pageledger compare-runs shows the per-page evidence for making it.

Tested Scale

The 0.1.x alpha has been tested locally on:

• 5,000 synthetic text pages (2.4s, ~2,100 pages/sec, artifact counts verified — raw files = provenance lines = quality lines = pages extracted).
• 1,000 synthetic text pages (0.25s, ~4,000 pages/sec).
• 72-page born-digital PDF (via pageledger[pdf]).
• 5 small checked-in text fixtures covering clean, multipage, OCR-noisy, blank, and short text.

This is evidence of the envelope, not a universal performance guarantee. Stress tests are marked @pytest.mark.stress and skipped in default CI. Run them with:

python -m pytest tests/pageledger/ -m stress

The Canonical Unit: Pages

PageLedger's defining decision is that the page is the canonical unit of work. It is the only unit every extraction backend shares:

• Cloud OCR (Textract, Azure Document Intelligence, Google Document AI, Mistral OCR) bills and reports per page.
• VLM/LLM extractors expose tokens — but only on model-backed paths; tokens are meaningless for classical OCR.
• Self-hosted engines (e.g. Docling) have no dollar cost at all, only compute time.

Because pages are the common denominator, routing, budgeting, and audit in pages lets you compare a Textract run against a Mistral run against a local model with one number. Tokens, compute-seconds, and dollars ride on top as optional, provider-conditional signals.

Every adapter reports a usage record where pages is required and everything else is optional:

usage = {
    "pages": 1,             # REQUIRED — the portable unit
    "tokens": None,         # VLM/LLM paths only
    "compute_seconds": None, # self-hosted engines
    "cost_usd": None,       # optional adapter-reported passthrough
}

Dollar cost is derived by PageLedger, never required of the adapter, in priority order: (1) adapter-reported cost_usd, (2) configured unit rates (cost_per_page / cost_per_1k_tokens), (3) otherwise null — the run still reports raw page counts. Budgets cap on pages, tokens, or dollars, whichever the config sets, because the page count is the only value always present.

Why Not Just Use OCR?

OCR and document-AI tools are improving quickly. Mistral OCR, Google Document AI, Azure Document Intelligence, AWS Bedrock Data Automation, Docling, Marker, Surya, olmOCR, OCR-D, and Unstructured all cover parts of extraction, conversion, layout analysis, validation, or workflow orchestration.

That makes PageLedger's scope deliberately narrow: it is useful when you need a local, provider-neutral ledger around those tools, not when you only need one tool to convert a document once.

Tool	Strong At	PageLedger Difference
Mistral OCR	Hosted OCR/document understanding with Markdown and table reconstruction.	PageLedger records why pages went to Mistral, what it cost, and which pages need review or rerun.
Google Document AI / Gemini layout parser	Managed OCR, layout parsing, table structure, and RAG-oriented chunking.	PageLedger keeps provider-neutral run artifacts and project-local audit queues outside a cloud workflow.
Azure Document Intelligence	Managed text, key-value, table, and field extraction.	PageLedger can track Azure runs beside local or other-provider runs with the same page-denominated manifest.
AWS Bedrock Data Automation	Managed classification, extraction, validation, HITL review, and business-rule workflows.	PageLedger stays lighter and filesystem-native for research projects that do not want a cloud stack.
Docling	Converting PDFs and documents into structured output.	PageLedger records which pages went to Docling, the usage and cost evidence, and which pages need review or rerun.
Marker	High-quality PDF-to-Markdown conversion.	PageLedger handles routing, run manifests, cost controls, and review queues around conversion.
Surya	OCR, layout analysis, table recognition.	PageLedger can call Surya as an extractor, then audit the result across a full archive run.
olmOCR	LLM-based PDF extraction with document-oriented output.	PageLedger adds page routing, quality diagnostics, cost evidence, and rerun manifests.
OCR-D	Mature OCR workflow model with METS/PAGE/ALTO conventions.	PageLedger aims to be lighter, VLM-aware, and friendlier to local research workflows that need JSON/YAML artifacts before full library infrastructure.
Unstructured	Document partitioning and preprocessing for downstream use.	PageLedger focuses on extraction run control: what ran, what passed, what failed, what cost money, and what should be rerun.

Non-goal: PageLedger is not an OCR engine, PDF converter, or layout detector. If a project already has clean structured text from a single tool and does not need routing, schema checks, review queues, cost tracking, or reruns, it likely does not need PageLedger.

For practical provider-agnostic choices, see docs/ocr-options.md. The recommended workflow: run cheap local extraction first, inspect quality.jsonl warnings to identify weak pages, re-extract just those pages with a stronger adapter via pageledger rerun runs/first/ --config stronger.yml --out runs/second/, then pageledger compare-runs runs/first/ runs/second/ to see what improved.

What PageLedger Cannot Do

PageLedger does not make OCR or VLM output correct by itself. It cannot calibrate confidence scores across unrelated extractors, guarantee text accuracy, infer a project's source citation, or make a right-to-left, mixed-script, tabular, or handwritten collection work without explicit adapter and schema configuration.

Its job is narrower: preserve enough evidence that a researcher can see what ran, what failed, what is uncertain, and what should be reviewed or rerun.

Why This Exists

Modern OCR and document-AI tools can produce impressive text, tables, and Markdown, but real archive work fails in smaller and more frustrating ways:

• Some pages are title pages, indexes, maps, appendices, blanks, or prose, not data pages.
• Expensive VLM calls should be routed deliberately instead of sprayed across every page.
• Model-emitted tables often have variant headers, shifted columns, missing rows, merged cells, or inconsistent numeric formats.
• Confidence scores are weak unless they are checked against schema, arithmetic, page totals, and model agreement.
• Long extraction runs need cost tracking, retry logic, checkpoints, and provenance that survives reruns.

PageLedger is for projects where "the model returned JSON" is only the beginning of the work. The current alpha is for proving local run evidence and adapter contracts before trusting larger extraction workflows.

Audience

• Digital humanities labs processing archival scans.
• Libraries and archives building reproducible extraction pipelines.
• Historians and social scientists extracting structured data from source books, statistical tables, reports, and registers.
• Civic, legal, nonprofit, and local-government document projects.
• Document-AI engineers who need a lightweight local control plane around Docling, Marker, Surya, olmOCR, Tesseract, or API VLMs.

Architecture and Roadmap

The alpha implements the run controller (budgets, retry, provenance, quality signals, audit queues, rerun execution, cross-run comparison) and the adapter protocol (thin wrappers that accept a routed page and return serializable content, warnings, confidence hints, and usage metadata — see docs/adapter-protocol.md).

Automatic page routing, schema alignment, audit grading, rerun execution, and staged CLI commands are design targets, documented with examples in docs/design.md.

CLI

Six commands ship in the alpha; run and rerun are the ones that extract:

pageledger run scans/ --config pageledger.yml --out runs/run-001/
pageledger run scans/ --config pageledger.yml --out runs/run-001/ --dry-run
pageledger run scans/ --config pageledger.yml --out runs/run-001/ --json --log-level INFO
pageledger rerun runs/run-001/ --config stronger.yml --out runs/run-002/
                                              # re-extract only the flagged pages
pageledger compare-runs runs/run-001/ runs/run-002/
                                              # page-by-page diff of two runs
pageledger init-config --out pageledger.yml   # write a minimal valid config
pageledger inspect-run runs/run-001/          # summarize a run directory
pageledger doctor --json                      # check optional dependencies

--dry-run generates the route map and planning artifacts without calling extractors, so you can inspect routing before spending money or trusting output. --json makes command status scriptable (parseable stdout, errors on stderr), and --log-level controls the detail written to the run log.

Three adapters ship built in. text writes each page of a UTF-8 text input to raw/ (pages split on the form-feed character). pdf_text reads born-digital PDF text layers through pypdf (install pageledger[pdf]). pdf_ocr OCRs scanned PDFs through locally installed pdftoppm and tesseract. For richer layout extraction, point run.adapter at a project adapter such as my_project.adapters:TesseractCliAdapter (use --adapter-path to make the module importable); PageLedger keeps the same manifest, cost, log, and provenance envelope around any engine.

The doctor output reports optional packages, common OCR commands, and whether cloud OCR/VLM keys are present without printing their values.

Starter config examples live in docs/examples/:

• pageledger.yml (recommended first-use combined config)
• page-taxonomy.yml
• table-schema.yml
• run-policy.yml

The split files are useful for documentation and larger projects. The recommended v0.1 user experience should start with one pageledger.yml that contains taxonomy, schema, and run sections.

Run Artifacts

A run should be inspectable without a database or service:

runs/run-001/
├── manifest.json
├── config-snapshot.yml
├── route-map.yml
├── raw/
│   └── doc_0001_page_0002.txt
├── normalized/
├── audit.json
├── audit.md
├── provenance.jsonl
├── quality.jsonl
├── cost.json
├── run.log
└── rerun-manifest.yml

The important artifact is not only the extracted data, but the evidence around it: which pages were skipped, which model or engine was used, what usage and cost evidence was observed, which pages failed or need review, and what should be rerun or reviewed.

manifest.json is the canonical run artifact. audit.md is a human rendering of audit.json, not a second independent source of truth. Quarantined pages live in audit.json under quarantine_queue. config-snapshot.yml preserves the exact user config that produced the run. run.log is JSONL — one line per extractor call with timestamp, page_id, adapter, status, and any error — so partial or failed runs stay greppable. Every artifact validates against the JSON Schemas in schemas/. The artifact directory keeps the name normalized/ for stability, but the current alpha does not write normalized records yet; schema alignment remains a documented design target.

Design Principles

• Record uncertainty; do not silently fix it.
• Treat heuristic confidence as evidence, not probability. Uncalibrated extractors should not imply certainty.
• Every run produces inspectable artifacts on disk.
• Adapters are thin: PageLedger does not own extraction, it owns the process around extraction.
• The manifest, route map, and provenance files should be useful without a running service or database.
• Preserve separate citations for software and source data.

Related Documents

• schemas/ — JSON Schema files, the machine-readable authority for artifact contracts
• docs/design.md — design targets: router, schema aligner, rerun policies, staged CLI, open research questions
• docs/comparison.md — positioning against the 2026 extraction ecosystem
• docs/run-manifest-spec.md, docs/route-map-spec.md, docs/audit-spec.md, docs/provenance-spec.md, docs/rerun-manifest-spec.md — per-artifact field specs
• docs/adapter-protocol.md — how to wrap an OCR/VLM tool as an adapter
• AGENTS.md — orientation for AI coding agents working in this repository
• CITATION.cff — how to cite PageLedger in research output