The Engine
Two pipelines, one findings schema, and a source for every claim.
Verdictums isn't a general-purpose model with a legal prompt in front of it. It's an evidence engine built for this job: one pipeline for documents, another for video and audio, both feeding a single schema in which every finding is reproducible, cited, and tracked in a custody log.
Pipeline A — Document AI
General-purpose OCR is trained on clean print and priced for volume. Legal evidence is rarely clean and never just volume. The document pipeline is built for water-damaged productions, cursive in the margins, third-generation photocopies, dense tables, and pages that have already been through someone else's redaction tool.
Stage 01
Intake & hashing
Each file is fingerprinted the moment it arrives. Its SHA-256 hash becomes the root of that file’s custody chain, before any processing starts.
Stage 02
Recognition
Degraded scans, handwriting, and awkward layouts go through models trained on exactly this kind of material rather than on clean, typed pages.
Stage 03
Structure extraction
Tables, forms, Bates stamps, and earlier redactions are detected and indexed so every element can be addressed and searched.
Stage 04
Findings & citations
Output lands in the unified schema. Each finding carries its page-and-line source, and the schema will not accept a finding without one.
Pipeline B — Video & Audio
An hour of bodycam footage is an hour of potential disclosure obligations. The media pipeline treats redaction, transcription, and attribution as one synchronized operation on one timeline, so nothing drifts out of alignment and every second can be accounted for.
Stage 01
Intake & sync
Bodycam, interview-room, and surveillance media are hashed and lined up on one synchronized timeline.
Stage 02
Redaction
Faces, license plates, and specified audio segments are redacted inside the pipeline rather than as a cleanup step afterward.
Stage 03
Transcription & attribution
Speech is transcribed and attributed to speakers across the whole timeline, including cross-talk and multi-party segments.
Stage 04
Findings & citations
Findings enter the same schema as documents, cited to the frame and second so they can go straight into the record.
The unified findings schema
Whatever the medium, output lands in one schema. A finding from a handwritten exhibit and a finding from an interview recording share the same shape: the claim, a confidence score, the citation, the model version that produced it, and a reference into the custody chain. Platforms integrate once and get every evidence type.
- ■One integration contract for documents and media
- ■Citations are structural, not optional metadata
- ■Custody references embedded in every finding
- ■Schema-stable across pinned model versions
finding_schema.json
{
"finding_id": "f-0047",
"claim": "Signature dated 14 Mar",
"confidence": 0.85,
"artifact": "exhibit_14b.pdf",
"citation": {
"type": "page_line",
"page": 14,
"line": 22
},
"model_pin": "vd-2026.01.3",
"custody_ref": "a94f...c21e"
}