probandum · Latin — “that which is to be proved”

AI legal analysis that cannot hallucinate.

Probandum builds a structured model of your matter. It knows the difference between an assertion and a fact — and shows its reasoning, traced to the source.

Now onboarding solo & small-firm attorneys for the pilot.

how an assertion becomes a “fact” — and how Probandum refuses

01 — beyond citation checkers

A complete legal reasoning engine.

Checkers confirm a citation exists. Probandum verifies the reasoning — catching the three failures that happen inside the model's memory:

claim laundering
A party's assertion retold as a “fact” — every later answer inherits the error.
first-draft thinking
The first plausible strategy wins; the alternatives are never examined.
sycophancy
Push back twice, and the analysis folds — agreeing with whoever argued last.

02 — the method

A checked model.

Provenance-typed memory

Every span keeps its origin: who said it, where, on what basis. Assertions stay assertions.

who said what, and whence

Deterministic verification

The model proposes; the graph disposes. Support and attack are computed by code.

computed, not asserted

Private by architecture

Documents are read on your machine. The cloud sees structure, not secrets.

raw data never leaves your machine

03 — the lineage

The chart Wigmore couldn't finish.

John Henry Wigmore, dean of Northwestern Law, in an undated portrait photograph — John Henry Wigmore (1863–1943). Harris & Ewing, Library of Congress — public domain.

In 1913, John Henry Wigmore — dean of Northwestern's law school — stopped trusting prose. A case is a mass of assertions, he argued, and prose is the worst vessel for it: smooth, persuasive, structurally dishonest.

“This article aspires to offer, though in tentative form only, a novum organum for the study of Judicial Evidence.” J.H. WIGMORE · THE PROBLEM OF PROOF · 8 ILL. L. REV. 77 (1913)

His fix: reduce every claim to a numbered symbol and chart the whole trial on one sheet. It worked — and vanished. A chart took weeks of handwork, so practice stayed with prose, and prose kept laundering assertions into facts.

Large language models change the arithmetic: a chart now takes minutes. But left alone, an LLM commits the very sin Wigmore charted against — it narrates, and narrative launders. So we use the model to draw the chart, and the chart to check the model.

The chart checks the model. Never the reverse.

Wigmore's 1913 evidence chart for Commonwealth v. Umilian — numbered circles and squares joined by inference arrows — J.H. Wigmore, “Evidence Chart for Commonwealth v. Umilian” (detail). Illinois Law Review, Vol. VIII No. 2, June 1913. Public domain — Google-digitized, Cornell University, via HathiTrust.

04 — see it argue

Same question. Two kinds of answer.

vanilla llm · chat

Did the defendant plan the murder?

Yes. Umilian planned to kill J. He had threatened revenge, and later sent the poisoned whiskey that did the job.

✗ stated as fact · no source · no distinction between assertion and proof

probandum · checked analysis

Did the defendant plan the murder?

“U. planned to kill J.” — the prosecution's theory, not a fact.

asserted-by-party status: unproven

support — threats (2, 5) · letter chain (10–17) · asserted
attack — no-motive hypothesis (18) · open
computed — plausible, contested · do not state as fact

A recorded demonstration is coming with the pilot — join the list to see it first.

05 — from the founder

Why I'm building this.

Andrzej Bednorz, founder of Probandum — Andrzej Bednorz · founder

I'm a second-time founder, a data scientist and a pro-se litigant. I've spent the past one and a half years studying how AI systems reason over evidence — and how quietly they go wrong.

Probandum is the system I need: it shows its trail of reasoning, keeps sensitive data local and never confuses confidence with proof.

— Andrzej Bednorz

06 — the list

Be first when the pilot opens.

Now onboarding solo & small-firm attorneys.