01 · Acquisition
The source comes before the case card
SimilarCase begins with public government decision text and approved public-record channels. Each source artifact is tracked separately from the logical decision so a corrected or reissued document can become a new source revision rather than an accidental duplicate.
- Official landing and document URLs are stored separately.
- Artifact and parsed-text hashes support reproducibility and change detection.
- Removed or replaced documents can be marked unavailable without erasing the audit history.
- Paywalled databases and third-party editorial headnotes are outside the acquisition policy.
02 · Parsing
Structure is recovered without losing the source order
Page-aware parsing separates procedural history, findings, reasoning, and disposition while preserving page and character context. OCR and layout recovery are used when ordinary text extraction is insufficient.
The structured case card keeps the original evidence alongside each derived field. This makes a factor inspectable rather than turning the model output into an unsupported summary.
03 · Extraction
Factors describe why the decision turned
Issues and outcomes
The question decided and the result on that specific issue.
Medical opinions
Role, weight, and the stated reason an opinion was accepted or rejected.
Determinative facts
Facts the reasoning treats as material to the disposition.
Legal application
The governing standard and its application to the record.
04 · Grounding
A field must lead back to evidence
Candidate quotations are checked against independently parsed source text. Grounding records preserve the exact passage, page when available, offsets, source identity, and verification state.
Fail closed
An unverified quotation cannot be promoted to a released case-card field. A technically extracted record may remain in review instead of appearing in the product.
05 · Retrieval
Ranking is multidimensional and inspectable
Search combines factor-aware semantic retrieval, lexical matching, issue overlap, structural filters, and authority signals. The active dimensions are weighted together, and result explanations expose the strongest matching passages and material differences.
Relevance scores help order a review set. They are not model confidence, legal authority, probability of success, or a substitute for reading the decision.
06 · Privacy boundary
User matters never become public corpus content
- Uploaded matter materials live in an authenticated, owner-scoped workspace.
- Dashboard, intake, result, profile, login, API, and shared-token routes are excluded from indexing.
- Claimant and medical details are not turned into public person profiles.
- Public-source corpus data and a user’s private claim file remain separate systems and purposes.