01 · Normalise
Atomic claims
Every statement in the testimony and exhibits becomes an atomic claim — entity · attribute · value — tagged with its source line and the speaker.
Litigation & forensic evidence analysis · South Africa
Datastroom IT normalises witness testimony and case documents into atomic, sourced claims — then surfaces where the record contradicts itself and which threads are worth pursuing. Every finding points back to the exact line it came from.
A working demonstrator runs the full pipeline on a live public inquiry.
The production line
01 · Normalise
Every statement in the testimony and exhibits becomes an atomic claim — entity · attribute · value — tagged with its source line and the speaker.
02 · Detect
The same fact asserted two ways. Conflicts are caught by a deterministic lookup that cannot invent one, plus model-assessed cross-referencing that stays cited and hedged.
03 · Surface
Contradictions, unreconciled money and missing links are ranked into leads — each with a line of inquiry and every claim cited to the record.
04 · Request
Each lead can auto-draft the PAIA request (Form 2) that compels the missing record — keyed to the lead, ready for an attorney to review and file.
Deterministic — lookup, cannot fabricate Model-assessed — cited & hedged
A worked example
Every artifact stays linked to the last — the dispute the contradiction opens, the lead that inherits it, the draft that inherits the lead. Illustrative example; the format is the real output.
Same tender, two irreconcilable values — caught by lookup on the shared entity and attribute.
Line of inquiry: obtain the signed tender agreement, invoices and proof of payment to establish the actual contract value and who authorised it.
Proven at scale
The pipeline was developed on the full public record of the Madlanga Commission. The same production line runs on a firm’s own matter — in private.
See the method run end-to-end on the demonstrator at whattheysaid.co.za.
How a trial works
Depositions, statements, transcripts, exhibits. Your material stays yours and private — nothing is published.
Claims extracted, contradictions detected, leads ranked — a deterministic core with model-assisted reasoning cited throughout.
A contradiction report and a ranked leads pack, each item linked to its source line, ready for a lawyer to act on.
Why it holds up
The strongest tier is a database-style lookup, not a model prompt. When it flags two statements as conflicting, they genuinely conflict on the same fact — the same guarantee as a database key violation.
Every contradiction cites the two source locations it joins. If a citation does not resolve to real text in your documents, the flag is dropped — never asserted.
Analysis runs on your matter in private. Nothing is published, and your material is never used to train a public model.
It finds internal inconsistency in the record — conflicting statements, figures that do not reconcile — not a verdict on who is truthful. That boundary is exactly what makes the deterministic tier dependable.
Start a trial
I’m looking for a South African firm to run a first trial with. Bring a matter you already know well and see what the production line pulls out of it.
Email info@datastroomit.comOr reply to the LinkedIn post that brought you here.