Published · Updated
The evidence problem is a structure problem
Evidence arrives unstructured: a scanned passport, a stack of medical records, a contract with three amended versions, an audio deposition, a filing with exhibits attached. Firms don't lose cases because they lack the documents — they lose hours re-reading them, and lose arguments because two versions of the truth were never compared.
Evidence intelligence inverts the workflow. Instead of humans reading documents into structure, the system reads them into structure once — extracted, typed, connected and searchable — and humans review what it found.
What happens when you upload
The pipeline runs in five stages: upload, extract, identify, score and connect. Any document — passport, contract, medical record, court filing — or an audio recording can enter it.
- OCR on scanned images, so paper and photos become machine-readable text.
- Meeting transcription with speaker diarization, so recordings become attributable statements.
- Entity extraction across 8+ entity types — names, dates, amounts, addresses, identifiers.
- Document classification across 30+ document types, so the system knows a filing from a lease.
- Legal relevance scoring, so the file's priorities surface first.
Contradiction detection
The stage that changes case strategy is comparison. When the same fact appears in multiple documents, evidence intelligence checks whether they agree: a date that differs between a letter and a filing, a party name spelled two ways, an amount that doesn't reconcile across invoices. Contradictions are flagged for attorney attention rather than discovered in a deposition or, worse, by opposing counsel.
Every extraction carries a confidence value, so reviewers know which fields to verify first instead of proofreading everything equally.
Evidence chains and the evidence graph
Facts don't matter in isolation; they matter as connections. Each extracted fact links back to its source document and forward to the entities it touches, building an evidence graph: 12 entity types tracked, relationships mapped between exhibits, and timelines auto-assembled from dated events. Because the graph spans your case library, cross-referencing isn't a manual task — the system connects a new exhibit to what you already have.
The output is a matter digest: documents summarized, contradictions listed, timeline rebuilt, parties resolved — with every claim traceable to the page it came from.
Scale, and where humans stay in
Batch processing handles up to 100 documents at once, so a production drop or a discovery upload becomes a processing job rather than a reading assignment. What it does not become: a decision. Extraction, flagging and linking run automatically; interpreting what a contradiction means for the case, deciding what's admissible and shaping argument remain with the attorney. The system surfaces the facts; the lawyer builds the case.
Frequently asked questions
What document types can evidence intelligence handle?
Over 30 document types — passports, contracts, medical records, court filings, invoices and more — plus audio recordings via transcription. Scanned and photographed documents go through built-in OCR first.
How does contradiction detection actually work?
The system extracts the same categories of facts — dates, parties, amounts, identifiers — across every document in the matter, then compares them. Where two sources disagree on the same fact, the conflict is flagged with both sources attached so the attorney can judge which is right.
Can I trace a fact back to its source?
Yes. Every extracted entity keeps its field-to-value-to-source mapping, so any value can be traced to the exact document — and page — it came from, with a confidence score attached.
Does the system replace a paralegal's document review?
It replaces the first pass: reading, classifying, extracting and cross-checking. Reviewers work from structured output — summaries, flagged contradictions, timeline — and spend their time on judgment calls rather than transcription.
Is evidence intelligence part of the AI workforce?
Yes — Vera, the evidence processor, handles extraction, classification, contradiction detection and evidence chains, feeding structured facts to Atlas for matter-level analysis. It's included on the Business plan.