White-label evidence engine / Verdictums
Verdictums reads the scans, handwriting, and bodycam footage that ordinary OCR gives up on, and returns findings a lawyer can actually cite: page and line for documents, frame and second for video. Same input, same output, every time.
2
Processing pipelines
Documents / Video & Audio
1
Findings schema
One contract for every media type
Pinned
Model versions
Same input, same output
Hashed
Every operation
Append-only custody log
00 / The Problem
A model that invents a source even once hands opposing counsel an easy afternoon. In evidence work there is no acceptable rate of fabrication. A finding either traces back to something real or it should not be in the output.
The big document-AI services are excellent on clean, typed pages. Give them a water-stained scan, cursive in the margins, or a third-generation photocopy with Bates stamps and the accuracy falls off right where cases tend to turn.
Most AI tools cannot tell you what touched a file, when, or with which model version. Without that record, even a correct result is open to procedural attack.
How it works
bodycam_0347.mp4
2.4 GB
Send degraded PDFs, scans, bodycam video, or interview audio through a single API call. We hash the file the moment it arrives, and that hash opens its chain of custody.
The pipelines enhance, transcribe, and extract using a pinned model version, so the same input produces the same output no matter when you run it.
Finding 12 of 47
Speaker 2 confirms location at 14:32
Each finding comes with its citation, a confidence score, and its custody hash, packaged so counsel can stand behind it when questioned.
01 / The Engine
Documents go through one pipeline, video and audio go through another, and both land in the same findings schema. Whatever the source, every claim in the output points back to where it came from.
Pipeline A
Built for the material general-purpose OCR struggles with: faded and damaged scans, handwriting, dense tables, Bates-stamped productions, and pages that have already been redacted once.
Pipeline B
Bodycam, interview room, and surveillance footage on a single synchronized timeline. Redaction, transcription, and speaker attribution happen in one pass, so nothing drifts out of alignment.
02 / Defensibility
Pillar I
Model versions are pinned, so running the same file twice gives you the same bytes twice. You can prove it by comparing hashes rather than asking anyone to take your word for it.
Pillar II
Every operation on every file is written to a hash-chained, append-only log, from the moment it arrives to the moment it is exported. If anything is altered, the chain breaks and you can see where.
Pillar III
Each finding carries its source: page and line for documents, frame and second for media. Counsel can put the original in front of the court without hunting for it.
Inside the platform
A REST API and one findings schema. Most teams have a working integration inside a sprint.
await verdictums.process({
artifact: "exhibit_a.pdf",
redact: ["faces", "plates"],
});_
Faces, plates, and voices redacted; every line attributed.
Speaker 1 / 00:14:12
Speaker 2 / 00:14:32
"I was at the north entrance the whole time."
Custody events are appended to the log as they happen and can be exported at any time.
Custody event
Artifact re-verified before export. Hash matches ingest record.
03 / Licensing
We don't sell to end users. Verdictums is licensed to the platforms they already use: e-discovery suites, public safety records systems, and FOIA processing tools. Partners integrate through our API and ship the capability under their own name.
api.yourbrand.com
█awaiting request
04 / The Pilot
Pick the evidence your current tools handle worst. We run it through Verdictums and through the document-AI services you already rely on, on identical inputs, and send you the scored comparison. If we don't come out ahead on your material, you owe us nothing.
Day 01
You choose the material. We hash every file on arrival and open a chain of custody before any processing starts.
Days 02–06
Verdictums and your incumbent services run on the same inputs. Every output is scored against criteria we agree on up front.
Day 07
You get a side-by-side report with cited findings, reproducibility proofs, and a concrete integration path if you want to go further.
05 / Who It Serves
Market 01
Review suites that process millions of pages per matter. Verdictums takes on the documents that stall a review: degraded productions, handwritten exhibits, and material that has already been redacted once, and returns structure and citations reviewers can rely on.
Market 02
Agencies with more bodycam and interview-room footage than staff hours to review it. The engine redacts faces, plates, and audio in the pipeline, transcribes with speaker attribution, and keeps every clip on one synchronized, citable timeline.
Market 03
Processors working against statutory deadlines with exemption obligations. Verdictums speeds up review, and its audit trail shows that each withholding and release decision followed a repeatable, documented process.
06 / In Practice
An hour of bodycam video comes back as a synchronized timeline: transcribed, attributed to speakers, redacted where the law requires it, and citable to the exact frame. Work that would normally take a paralegal days of scrubbing happens in a single pass, with a custody entry for each operation along the way.
07 / Questions
These are the things platform and procurement teams tend to ask first. For anything deeper, the pilot exists so the engine can answer on your own data.
Ask something elseNo. We license only to platforms and integrate through an API. Your users see your interface and your brand; the engine does its work behind the scenes.
Model versions are pinned for each engagement. Run the same file through the same pinned version and you get byte-identical output. That is something you can demonstrate to a court, an auditor, or opposing counsel by comparing hashes.
Each finding carries its source location: page and line for documents, frame and second for media. The output schema has no room for an uncited claim, so a finding without a source simply cannot be produced.
Pilot material goes through the same chain-of-custody controls as production work. Once the benchmark is done, the data is destroyed or returned on day seven, whichever the pilot agreement specifies.
Whatever you use today. Most pilots compare Verdictums against the major cloud document-AI services on identical inputs that you select, usually the material those services handle worst.
Verdictums is based in Singapore and was founded in 2026 by Charles Armin. You can read more on the About page or reach the team directly at contact@verdictums.com.