Research that hides from its strongest critics is marketing. So the objections get their own page, stated at full strength, answered seriously — and where an objection is simply right, we say so.
Partly conceded. True, and stated on our front page before any critic said it. The difference we argue for is not purity but auditability: a model's bias can be measured, published, and corrected in the open, and the same inputs give the same outputs — so the bias is at least visible and fixable. A human judge's bias is invisible, unmeasurable, and dies with his career. You cannot audit a judge's heart. Choose which failure mode you'd rather live with — that choice is the research question.
Conceded — this is the strongest technical objection, and it's ours as much as the critics'. A thousand flawed human judges at least fail in a thousand different directions; one flawed model fails identically everywhere, at scale, before anyone notices. Any real deployment would need versioned models, published error rates, mandatory human appeal, and the failures published prominently — which is exactly the discipline this research programme binds itself to. If correlated error turns out to be unmanageable, that finding gets published as loudly as anything else, and the ladder stops at the rung the evidence supports.
They did, and their families are alive and grieving still — that is why the rules exist. We rule only on concluded cases with complete public records; the analysis is of judicial reasoning, never content about victims; and any case still before a court is out of bounds absolutely. But note what "tasteless" quietly asks for: that the most consequential decisions society makes — verdicts — stay beyond public re-examination forever. Every one of these trials was broadcast to millions as it happened; studying the record afterwards, with named sources and open methods, is the respectful version of the world's attention, not the exploitative one.
No — and this refusal is the programme's spine, not its cowardice. A machine that rules on half-heard evidence is doing exactly what the manifesto condemns in humans: judging from headlines. Verdicts on the living, mid-trial, could prejudice real proceedings and brand real people — research doesn't get to do that, and won't here. When a record is complete and the courts are done, a case becomes eligible for the docket. Until then, we analyse process, never people. The discipline is only worth something if it holds when it's tempting.
Taken seriously — this may be the deepest one. There are moments in judging that aren't rule-application at all: mercy for a first offender, a community speaking through a human peer. And a verdict people won't accept is not justice, however sound its reasoning. This is why the research ladder confines machine adjudication to low-stakes, high-volume matters — and why humans hold the gavel above every rung. Whether legitimacy can ever transfer to a machine is a cultural question no benchmark settles; our research measures what can be measured and says so plainly about what can't.
Sharpest of all, and half right. Somebody trains the model, updates it, decides what counts as evidence — power doesn't vanish, it relocates. The honest claim is therefore not "untouchable" but "visible": every version of the rules on public record, every judgment carrying its full reasoning, tampering detectable because the machine's behaviour is reproducible and the record is open. A corrupted judge can hide for a career. A corrupted model contradicts its own published record the moment it's re-run. That's not incorruptibility — it's corruption with an audit trail, which no human bench has ever offered.
Have a stronger objection? Send it through the form on the front page. The best ones get added here, at full strength, with our answer — or our concession.
← Back to the research question · The docket · Research note 01