What the research is working on right now, what's queued, and what's under consideration. This page is the honest state of the programme — it moves when the work moves.
Status as at 31 August 2026
The most famous verdict on earth — and the perfect first experiment, because the system itself returned two opposite answers on the same facts: a criminal acquittal, then civil liability. The complete public record survives: over half a million transcript lines, 3,900+ catalogued evidence items, and the courtroom video, all publicly archived.
And the most important thing the record does not contain: the jury's reasoning. Four hours of deliberation, no transcript, no record — the black box in its purest form. The machine's verdict will be published next to both human verdicts, with every step of the reasoning open.
Status: public record catalogued and verified · extraction scope being drafted · glass-box format in design.
The South African counterweight to Case 001 — and in one way the deeper experiment. Where the Simpson jury left no reasoning at all, Pistorius was tried by a judge who published her full reasons — and a higher court then examined that reasoning and found the law misapplied. Three layers of written human reasoning for the machine's analysis to stand beside: the trial judgment, the 2015 appeal, the 2017 sentencing appeal. The pairing is the study: the jury nobody could review, and the judge somebody did.
Status: queued behind Case 001 · official judgments verified on SAFLII · testimony record heavier to assemble (televised, but no structured transcript archive exists).
A lawyer from a courthouse dynasty, convicted of murdering his wife and son on circumstantial and digital evidence. The research question: how does a machine weigh a case with no eyewitness and no weapon — and how does its treatment of the cellphone timeline compare with the jury's? Post-trial challenges to the verdict form part of the assessment before any run.
The Theranos fraud trial — a case decided almost entirely on documents and intent. The research question: white-collar cases turn on what someone knew and when; how does a machine's reading of intent from a paper trail compare with twelve jurors' reading of a face on the stand?
A conviction resting substantially on the credibility of accusers testifying about events decades old. The research question: credibility assessment is the thing humans claim as their unique strength — what does a machine's reasoning even look like there, and where must it honestly say "this a machine cannot weigh"?
Two courts, overlapping factual questions, opposite outcomes: a London judge ruled against Depp on the evidence; a Virginia jury, watched by the entire internet, found for him. The research question: same facts, judge versus jury, media pressure at maximum — the closest thing to a natural experiment the modern record offers.
One killing of a law professor; four defendants tried before four different juries across six years, with a hung jury and a retrial along the way. The research question: run the same machine across all four records and measure consistency — the thing the human system visibly struggled to deliver here.
The same case tried twice — a hung jury, then an acquittal — with both full records preserved. As consistency research, gold. And a standing rule applies at full strength: the defendant is living and was acquitted, so the machine's published analysis is scoped to the process and the two juries' divergence — Judgir will never publish a verdict on her that goes beyond what the courts found. The case study here is the system, not the person.
The original "trial of the century," decided on handwriting, wood forensics and circumstantial chains — and argued about by historians ever since. The research question: what does a modern machine make of 1935 forensic standards, and does the reasoning that convicted Hauptmann survive a glass box built ninety years later?
The second South African case: a family murdered with an axe, a survivor convicted on a fully reasoned written judgment, appellate courts declining to disturb it. Together with Pistorius it gives the research two judge-tried SA cases with published reasoning — the richest possible comparison material for a machine that must also show its work.
Cases are chosen for the quality and completeness of their public record and the distinct research question each poses — jury opacity (001), corrected judicial reasoning (002, 010), circumstantial and digital evidence (003), documentary intent (004), credibility (005), judge versus jury on the same facts (006), consistency across juries (007, 008), and historical forensic standards (009). Never for how clever the machine will look. Each case passes a record-and-appellate-status assessment before any analysis runs, and the docket's statuses update as that work happens.
← Back to the research question · Research note 01 — the landscape