News / public record

Checking
  1. Update / Component / Scope · CheckingOLC Engine: a renewed presentation of legal workThe bilingual Engine page now presents the approved version 0.1.0 scope, its legal-work method and the boundary between current capabilities and planned expansion.
  2. Article · CheckingA Server in the Office Is Not a Security StrategyWhy an office server, cloud service, email route or encrypted backup must be assessed through capabilities, responsibility and evidence—not location alone.
  3. Article · CheckingWho Controls the Legal AI Stack?Public code can enable inspection without proving control of the whole system. Follow versions, provenance, data sovereignty, exit and regulatory roles in legal AI.
  4. Article · CheckingConfidentiality Does Not End at the Chat WindowConfidentiality is a property of the complete data path, not the chat window. What do DPAs, de-identification and local deployment solve — and why is the OLC gateway still only a concept?
  5. Article · CheckingWhen a High Score Hides a Critical Legal ErrorA high benchmark score does not establish that legal AI is dependable for every problem. This article explains how to measure critical errors, retrieval, abstention, evaluators and the path to a legal answer.
  6. Update / Component / Readiness · CheckingWord Connector public presentation is completeThe Apache-2.0 v0.1.0-beta.1 source-only release now has a complete public component record, homepage feature, Roadmap outcome and discovery path.
  7. Update / Component / Release · CheckingOpenLegalCore Word Connector v0.1.0-beta.1 is publicThe Apache-2.0 source-only beta provides a verified Word for the web path for researching Slovenian legal sources through OLC Engine, directly or through Open WebUI.
  8. Article · CheckingLegal RAG Can Fail Before the Model AnswersA genuine provision can still support the wrong answer. This article shows how collection scope, document splitting, filters and ranking determine which sources a legal AI model ever sees.
  9. Article · CheckingWhen a Correct Citation Leads to the Wrong LawA genuine provision and an official link do not prove that the right law was applied. This article explains why legal AI must preserve temporal versions, transitional rules and the source-selection path.
  10. Update / Project · CheckingPublic project foundation gate is completeOpenLegalCore has completed its public project foundation gate: core public routes, bounded participation channels and publication controls are operating.
  11. Article · CheckingAn Audit Trail Is Not a Log: What a Digital Legal Process Must PreserveAn audit trail is useful when it permits a reasonable reconstruction of the particular process while remaining purpose-bound, appropriately protected and subject to retention rules.
  12. Article · CheckingArticle 86 of the AI Act: Does the right to explanation already apply to Annex III systems?The legal question remains open. The technical capacity to provide an intelligible explanation should not.
  13. Update / Project · CheckingPrivate security reporting is now availableOpenLegalCore now provides a monitored project-wide private channel and a canonical security-reporting policy.
  14. Update / Component / Release · CheckingSlovenian Case Law Pipeline v0.1.7 is publicOpenLegalCore has published the production-verified Slovenian case-law ingestion component as a source-available BUSL-1.1 release.
  15. Article · CheckingA Result Is Not a Method: Why Visible Methods Matter in Legal AIWhy legal AI needs inspectable sources, provenance and human review: a legal and technical analysis of the EU Artificial Intelligence Act and GDPR.
  16. Update / Component / Release · CheckingSlovenian Legislation Pipeline v0.1.0 is publicOpenLegalCore has published the production-verified PISRS legislation-ingest component as a source-available BUSL-1.1 release.
  17. Update / Component / Release · CheckingLegal OCR Pipeline v0.1.2 is publicThe first public OpenLegalCore component is available with code, tests, offline review tooling and a bounded acceptance record.
Explore the code
Menu
Component register

COMPONENT 01 / PUBLIC · v0.1.2

Legal OCR Pipeline

A command-line OCR application. A reusable system component.

It converts scanned and image-based PDFs into ordered text using Google Vertex AI. Use it on its own or adapt it for a larger document workflow.

Inspect the acceptance record Open the public repository
Status
PUBLIC
Version
0.1.2
Licence
APACHE-2.0
Verified
2026-08-22
  1. 01

    AI-native recognition

    Printed text, stamps and handwriting.

  2. 02

    Deterministic controls

    Page order, validation, retries and checkpoints.

  3. 03

    Human verification

    Offline source-and-transcript review.

  4. 04

    Standalone or embedded

    Standalone command-line use; adaptable for integration into other systems.

01 / CAPABILITIES

AI recognition inside a controlled pipeline.

The model performs probabilistic visual recognition. The pipeline defines rendering, page identification and order, validation of page identifiers and response completeness, retries, and the saving of checkpoints and final output. These steps are inspectable, but this does not guarantee identical transcripts across runs. The accepted campaign used one difficult legal filing. The architecture is reusable across image-document workflows, but no cross-domain performance benchmark is claimed.

  1. 01PDF
  2. 02Page render
  3. 03Multimodal AI
  4. 04Validation
  5. 05Checkpoint
  6. 06Ordered artefacts
  7. 07Human review

INPUT

Documents that need visual reading.

  • Image-based and scanned PDFs
  • Page images with printed and graphic text, stamps or handwriting
  • One PDF per run; at most 180 pages in the current configuration

OUTPUT

Artefacts that can travel through a larger system.

  • Ordered ocr.txt document output
  • Page checkpoints and a machine-readable metrics.json record
  • Campaign metrics and review-ready source/transcript pairs

STANDALONE USE + INTEGRATION PATTERNS

  1. 01Standalone command-line application
  2. 02Document processing is included in Engine 0.1.0
  3. 03Standalone release and Engine have separate records
  4. 04Possible integration with a custom document workflow

02 / CONTROLLED PROCESS

The model reads.The pipeline controls.

The model recognises text; the pipeline controls processing and preserves intermediate results. Automated checks do not establish transcription accuracy: the transcript must still be checked against the source.

  1. 01

    Render

    The pipeline converts the PDF into page-identified JPEG images at a controlled resolution and quality.

  2. 02

    Recognise

    It sends each page to a multimodal AI model for recognition of printed and graphic text, stamps and handwriting.

  3. 03

    Validate

    It checks page identity and the terminal control marker before accepting the response.

  4. 04

    Checkpoint

    It saves completed pages so processing can resume. Pages whose responses have already passed automated validation do not need to be sent to the provider again.

  5. 05

    Assemble

    It writes the combined text in document order only after every page succeeds. Checkpoints and machine-readable metrics are retained.

  6. 06

    Review

    A person compares source pages with their transcripts locally using the public offline review tool. Keep the generated review package private.

03 / HUMAN VERIFICATION

Review stays beside the source.

The repository's public offline review tool displayed each original page beside its OCR transcript. The generated review package remained private, and review ran locally with no network or provider calls.

  1. 01

    Deterministic sample

    A mandatory sample combines fixed controls, no-text pages, edge cases and pseudo-random pages. Its exact page list is public in the canonical source record.

  2. 02

    Offline review

    Review runs locally without network or provider calls. Source pages, OCR output and generated review artefacts remain private.

  3. 03

    Explicit classification

    Every reviewed page is classified as PASS, MINOR, ISSUE or UNCERTAIN; the aggregate status is recorded separately.

RECORDED ACCEPTANCE

How the result became evidence.

The result must be read together with the document scope, execution record, reproducible sample selection, human review and stated limitations.

  1. INPUT BOUNDARY

    One private source, fixed rendering.

    One 174-page filing was rendered at 150 DPI as JPEG · quality 95. The source and generated acceptance artefacts remain private.

  2. EXECUTION RECORD

    The complete CLI path was measured.

    174 page calls used Gemini 3.5 Flash-Lite on Google Vertex AI, with up to 16 concurrent page tasks. The record shows 0 retries and 0 failed pages.

  3. REVIEW RULE

    The sample was selected deterministically.

    The mandatory 40-page sample combined fixed boundary and known-interest pages, all no-text pages, length extremes, illegible markers and a fixed-seed selection.

  4. CLAIM BOUNDARY

    The record stops where the evidence stops.

    SAMPLE_PASS records one bounded campaign. It is not a universal accuracy claim, character- or word-error rate, or service-level promise.

PUBLIC RECORD

Method · configuration · sample rule · aggregate result

PRIVATE MATERIAL

Source filing · OCR output · metrics and review artefacts

Inspect the complete method description

04 / BOUNDARIES + NEXT GATES

Where the evidence stops.

The acceptance supports one working path on one difficult filing. It does not turn model output into an authoritative legal source or a universal performance promise.

  1. 01

    This is one concrete 174-page legal filing and one acceptance run.

  2. 02

    The timing is not an average of repeated runs and is not a guarantee for other documents.

  3. 03

    The review result is SAMPLE_PASS, not FULL_PASS: 40 of 174 pages were manually reviewed.

  4. 04

    OCR is not an authoritative legal source. Critical names, dates, amounts, deadlines, quotations and operative text must be checked against the source document.

  5. 05

    The current fixed 180-call campaign cap means one new run cannot accept more than 180 pages without changing the configuration.

  6. 06

    Document pages are sent to Google Vertex AI. The component makes no independent data-residency, confidentiality, regulatory-compliance or provider-retention guarantee.

  7. 07

    The publicly released package produces ordered plain text with marked page boundaries and separate checkpoints. It does not preserve complete page layout or extract structured legal facts.

  8. 08

    The package does not provide a hosted service, packaged library API, container image, graphical application or dedicated Open WebUI connector.

Inspect every source-record limitation

READINESS + ROADMAP

Published release. Further accuracy and speed evaluation.

Later work remains gated by measured quality. A faster provider path is not adopted until it preserves the required recognition reliability.

  1. 01

    PUBLIC BASELINE

    v0.1.2 is published under APACHE-2.0.

    The public repository includes the component code, documentation, tests, offline review tooling, security policy and canonical acceptance record.

  2. 02

    ACCURACY

    Broaden the evaluation corpus.

    Add difficult layouts, handwriting, stamps and multiple document families while keeping review outcomes and error classes explicit.

  3. 03

    LATENCY

    Measure and optimise every stage.

    Separate render, provider, validation, checkpoint and artefact-write latency before making further public speed claims.

  4. 04

    PORTABILITY

    Evaluate additional model paths.

    An experimental Cerebras path has already been tested and demonstrated a substantial latency advantage, but it did not meet the required OCR reliability threshold. It is not part of v0.1.2 and will be re-evaluated as stronger models become available.

    EXPERIMENTAL / QUALITY GATE NOT MET

05 / RELEASE RECORD

A public building block with explicit boundaries.

Version 0.1.2 is public and useful by itself or embeddable inside larger systems. The software package is open; the confidential acceptance source and generated campaign artefacts remain outside the public repository.

RELEASE IDENTITY

Source, status and accountability.

Component
Legal OCR Pipeline
Version
0.1.2
Status
PUBLIC
Acceptance record
RECORDED ACCEPTANCE ↗

RUNTIME + OUTPUTS

What the package operates.

Runtime
PYTHON 3.10+ · POPPLER · VERTEX AI / ADC
Command
python vertex_parallel_pdf_ocr.py input.pdf output-directory
Outputs
checkpoints/ · metrics.json · ocr.txt
Public package
CODE · DOCUMENTATION · TESTS · OFFLINE REVIEW TOOLING

EXCLUSIONS + PRIVATE BOUNDARIES

What remains outside.

Service boundary
NOT INCLUDED: HOSTED SERVICE · LIBRARY API · CONTAINER IMAGE · GRAPHICAL APPLICATION · DEDICATED OPEN WEBUI CONNECTOR
Data flow
DOCUMENT PAGES ARE SENT TO GOOGLE VERTEX AI
Private boundary
SOURCE FILING · OCR OUTPUT · METRICS · REVIEW ARTEFACTS · CREDENTIALS
Brand rights
SEPARATE FROM THE CODE LICENCE