A new foundationfor litigation.
Crimson's proprietary ingestion and retrieval architecture is a major reason it outperforms general-purpose AI on litigation. At its center is the Case Model: a rich digital representation of the matter, which Crimson builds at upload and maintains for the life of the dispute.
The structure
behind the strategy.
This is the Case Model for a warranty claim after an acquisition. Hover any issue, fact, event, or entity to follow its connections through the record.
Northbridge v. RedwoodCase Model“We need to raise the continuing delivery delays before month end.”
Kate Preston·Email to NovusGrid management · Mar 14, 2022p. 1
0
Facts traced
0
Events traced
0
Entities traced
Why general-purpose AI
falls short in litigation.
Most AI tools treat a case file as a collection of text to search when a question is asked. But a case is not just text. It is a structure of claims, defenses, issues, and evidence, interpreted in context and at a particular stage of the proceedings.
Understanding of the matter
General-purpose AI
Rebuilt from scratch with every prompt.
Crimson
A persistent understanding of the whole matter, built at upload and updated as the case evolves. It does not reset with each question.
What it finds
General-purpose AI
Passages that sound like the question. Key evidence phrased differently is easily missed.
Crimson
The facts, events, and evidence connected to the issue, even in documents that never use the same words.
Coverage across the case file
General-purpose AI
Misses more as the case file grows, leaving you unsure what has been overlooked.
Crimson
Consistently high recall across the whole case file. Nothing material slips through.
Context
General-purpose AI
Every passage looks the same. A denial in a pleading can be mistaken for an admission.
Crimson
Reads each passage in light of who wrote it, where it appears, and the stage of the case.
Provenance
General-purpose AI
Vague or missing citations that leave every answer to be checked by hand.
Crimson
Every fact and inference cited to document, page, and passage. Verifiable in seconds.
The hard work is done
before you ask.
At upload, Crimson performs complex, compute-intensive processing to turn the case file into the Case Model. We encode deep domain expertise into the pipeline, so Crimson understands not only what information says, but how to interpret it given where it appears and the stage of the case at which it arises.
Building the Case ModelNorthbridge v. Redwood- 01
Upload
Documents arrive from your DMS, inboxes, and data rooms.
- RE_ Delivery delays (K Preston).msg48 KB
- Board minutes 05.19.22.pdf212 KB
- SPA (executed).pdf1.8 MB
- Disclosure schedules.pdf640 KB
- Mercer Decl. (draft 3).docx96 KB
- Notice of claim.pdf120 KB
0 documents
- 02
Classify
Crimson categorizes, OCRs, and de-duplicates every document.
Correspondence0Board papers0Contracts0Pleadings0Witness and expert00 duplicates removed
- 03
Extract
Crimson pulls out events, parties, dates, and defined terms.
0
Events
0
Parties
0
Dates
0
Defined terms
0 across the record
- 04
Interpret
Crimson reads every passage in its full context.
Email · K. Preston · Mar 14, 2022
“We need to raise the continuing delivery delays before month end.”
- IssueSellers' knowledge
- Timing108 days before approval
- AuthorHelix, key customer
0 passages read in context
- 05
Link
Everything connects into the Case Model, cited to the paragraph.
0 links in the Case Model
A purpose-built
legal ontology.
The Case Model maps the relationships between nodes that describe how a dispute is argued and decided.
Evidence
Documents, passages, exhibits, and testimony, each identified to page and passage.
Events
Dated occurrences extracted from the record and placed on a single, reconciled chronology.
Parties
Claimants, defendants, witnesses, experts, related entities, and their roles over time.
Disputed facts
Propositions asserted by one side and contested by the other, with the evidence each relies on.
Claims
Causes of action, heads of loss, and the facts and propositions each depends on.
Defences
Denials, limitation, set-off, and affirmative defenses, mapped to the claims they meet.
Issues
The questions the tribunal must decide, linking claims, defenses, facts, and evidence.
Legal propositions
Legal rules, contractual terms, and standards that determine how facts are interpreted.
Continuously evolving.
Fully auditable.
As your team uploads new material and works with Crimson, the Case Model continuously evolves while preserving full provenance. You can trace every fact, inference and relationship back to the precise supporting document, page, and passage.
The result is an evidence-backed intelligence layer purpose-built for high-stakes litigation: one you can rely on in front of a tribunal, because you can check every step of the reasoning.
Disputed factThe board knew of the Helix delays before approving the acquisition.
Issue 1
Did the sellers know about the Helix delays?
Source
Board minutes, NovusGrid, Inc. · May 19, 2022 · item 3
Passage
“Ms. Caldwell reported that Helix had raised continuing delays with deliveries and that a remediation plan was in hand.”
Multiple retrieval strategies,
orchestrated by agents.
When a user asks Crimson a question, the platform draws on the Case Model alongside full-text search, embedding-based retrieval and other techniques, orchestrated through a multi-agent architecture. Because Crimson has already understood and structured the matter, it retrieves relevant information across extremely large case files with exceptionally high recall.
Case Model traversal
Agents navigate the ontology directly: from an issue to the disputed facts beneath it, to every piece of evidence bearing on each fact.
Full-text search
Exact-match retrieval across the entire record for names, dates, clause references, and defined terms.
Embedding retrieval
Semantic search for passages that express the same idea in different words, across every document type.
Multi-agent orchestration
A planning agent breaks down the question, delegates to specialist retrieval agents, and reconciles the results into a single cited answer.
Most AI tools are like an associate who was handed the case file this morning. Crimson is like one who has lived with the case for a year.
The first associate
May identify some relevant documents, passages, or phrases, but will struggle to be comprehensive. No established mental model of the case, no understanding of how the pieces fit together, and no systematic way of surfacing important but less obvious evidence.
The second associate
Has a deep understanding of the matter, knows how the evidence fits together, has systems for locating relevant information and an established mental map of the case. Ask them a question and they reason across the matter, identifying the relevant evidence, even when it is distributed across thousands of documents.
Questions from technology and knowledge teams.
Further readingWhy RAG alone isn't enough for litigationA vector database stores passages by similarity and a generic knowledge graph stores entities and relationships. The Case Model is a legal ontology: its node types and relationships are the concepts litigators actually reason with. Claims depend on facts, evidence supports or contradicts facts, issues connect claims and defenses, and propositions determine how to interpret facts. Crimson uses embeddings and full-text search as well, but orchestrates them through this structure rather than using them in its place.