How Defacta works: an independent verification layer for documents, archives and AI
By Galia · Published
Archives keep growing. Reports, filings and news flow in faster than anyone can check them. Most teams no longer have a writing problem. They have a checking problem.
Sounds familiar?
- Analysts: AI-assisted drafts look finished, but checking every figure and source takes longer than writing them.
- Legal and compliance: The citation is real, but does it support the claim? And who verified it?
- AI teams: Knowledge bases mix current and outdated information, sometimes with hidden instructions targeting AI agents.
- Publishers and platforms: Submissions keep growing, but review capacity doesn’t.
Documents often tell different parts of the same story. Defacta connects the dots across them, uncovering overlaps, dependencies and contradictions, while making the information easier to explore, analyse and reuse in AI applications.
Defacta turns documents, archives and incoming information flows into a verifiable foundation for decisions, knowledge bases and your own LLM.
How Defacta works
Defacta combines data structuring, document intelligence and verification analysis in one workflow. A document or archive is prepared once. Verification and other tasks then run on the same structured data.
1. Document preparation (done once)
The pipeline turns texts, multi-page reports and whole archives into clean, structured data. It goes beyond OCR: text recognition is combined with parsing of layout, reading order, tables, headings and figures.
It produces two formats:
- Markdown — clean text with headings and tables, free of layout noise, ready for language models.
- JSON — preserves structure. Each element’s type, page and coordinates are kept so every verification result can be linked back to its exact place in the original.
Prepared documents can be exported in the format needed for language models, AI agents, knowledge bases, RAG pipelines, or as reports for people. Once prepared, the same data can be reused for new tasks without reprocessing.
2. Verification and tasks on prepared data
On the structured material, Defacta runs the checks you need. The core task is claim verification across four layers:
- Factual verification — Claims are checked against credible sources, authoritative databases or your own knowledge base. Hallucinations, fabricated citations and contradictions between text and tables are detected and flagged.
- Source credibility — Every source is evaluated for reliability, track record and relevance. Anonymous or low-credibility sources are identified.
- Bias and manipulation — Manipulative framing, selective omission and persuasion tactics are detected. Bias is scored across dimensions and shown without passing judgement.
- Misinformation registry — Claims are cross-referenced against published fact-checks and known false claims (integrations in progress).
How verification is performed:
- A model ensemble is selected for each analysis. Disagreement between independent models is treated as a signal.
- Sources include fresh search, trusted references, Defacta registries, your own knowledge base, and partner sources.
- The process is configurable: web sources only, internal knowledge base only, or both. Sources, rules and depth of analysis can be set to the task.
- Contested, low-confidence and high-risk cases go to human review. AI accelerates individual steps; people decide.
Additional tasks on the same prepared data include:
- Finding contradictions between text and tables
- Preparing a verified corpus for RAG and knowledge bases
- Checking incoming documents before an AI agent acts on them
- Detecting hidden instructions in documents
What you receive
- A clear report: short plain-language summary followed by layer-by-layer findings, sources, risks, limitations and review context
- Claim labels (confirmed, unconfirmed, contradicted, outdated) linked to sources and, where structured, to exact pages
- Prepared documents ready for export (Markdown and JSON for LLMs, agents, knowledge bases and RAG; PDF, DOCX, XLSX and TXT for other uses)
- Full audit trail: job history, source metadata, timestamped SHA-256 hashes, integrity checks and W3C PROV provenance
- Revision workflow: edit text, generate repair drafts, re-check changes and compare versions
Designed for real workflows
For people and decisions
Analytical reports, pre-publication checks, documents before filing, pre-moderation, and personal or professional verification of texts, articles, letters or AI answers.
For AI systems
Checking archives before indexing for RAG, cleaning datasets for fine-tuning, verifying incoming documents before agents act on them, detecting prompt-injection style hidden instructions, and filtering external sources before they reach a model.
Defacta is deliberately not a verdict machine. It shows the evidence. Bias is surfaced, not judged. Authorship (human or AI) is irrelevant — only the content is checked. Every result is reproducible through the audit trail. The system is designed to be neutral; systematic bias, if found, is investigated and corrected.
Availability
Available now: text verification in six modes, four-layer analysis (registry integrations excepted), model ensemble, reports, audit trail, revision workflow, exports, browser extension and shareable reports.
In pilot: structuring of multi-page documents and archives, claim labels with page links, prepared data for RAG and AI, checks against an organisation’s own knowledge base.
In development: API and client integrations, broader misinformation registry connections, MCP support, and specialised models for annotation and claim extraction.
Defacta does not issue verdicts. It shows the evidence. People decide.