Citation, not confidence.

There's a pattern in how AI failures reach the legal press. A lawyer submits work containing citations to cases that don't exist. The tool that produced them wasn't malfunctioning — it was doing exactly what it was designed to do: generate plausible text. The failure wasn't the hallucination. The failure was that the output sounded right, and sounding right was all the system was ever optimised for.

That's the confidence problem. And it's why we built Pivot the way we did.

Confidence is the product of probabilistic systems

A large language model, asked a question, produces the most statistically likely response. It doesn't know whether the response is true. It doesn't distinguish between something it read in your documents and something it pattern-matched from training data. And critically, its fluency is constant: the wrong answer arrives in the same assured tone as the right one.

For many tasks, that's fine. For drafting a first pass of an email, plausibility is the point.

Legal evidence is not one of those tasks. When a tool tells a solicitor that the parties made conflicting claims about a pension, one of two things is true: the conflict exists in the disclosure, or it doesn't. There is no version of that question where a plausible answer is acceptable. And in England and Wales, the responsibility for the answer sits with the solicitor — not the tool. Courts and regulators have been unambiguous about that.

So the question every firm should ask of a legal AI tool isn't "how accurate is it?" It's simpler: when this system gives me an answer, can it show me where the answer came from — and if it can't, does it stay silent?

What deterministic looks like in practice

Pivot doesn't generate answers about your case. It builds your case as a structure and reads answers off that structure.

When disclosure comes in — Form E, bank statements, valuations, message exports — Pivot extracts every claim it contains: this party stated this asset at this value, in this document, at this page and paragraph. Those claims become nodes in a graph. Where two claims about the same asset disagree, the graph contains a contradiction edge between them.

Now consider what happens when a solicitor asks: are there any conflicting valuations in this case?

A probabilistic system would predict an answer. Pivot traverses the graph. Either a contradiction edge exists — in which case the answer comes back with both claims, both sources, and the exact citations — or it doesn't, in which case the answer is no. The system is structurally incapable of inventing a conflict, because a conflict isn't a sentence it generates. It's an object that either exists in the graph or does not.

The same architecture governs everything Pivot surfaces. A gap in disclosure is an expected node with no evidence attached. A Section 25 analysis is a mapping from statutory criteria to the evidence nodes that support them. Every output has a provenance trail back to source, because the output is the trail. There is no mode in which Pivot asserts something it cannot cite — not as a policy, but as a consequence of how the system is built.

Why this is an architecture question, not a testing question

Some vendors address hallucination with guardrails: a second model checking the first, confidence thresholds, disclaimers. These reduce error rates. They don't change the underlying fact that the system is generating text and hoping it corresponds to reality.

The distinction matters because legal work is adversarial and audited. Opposing counsel will check your citations. The court may examine how a conclusion was reached. A tool whose accuracy is statistical — right 97% of the time, with no way of knowing which 3% is wrong — puts the burden of verification entirely back on the solicitor, which erases most of the value the tool was supposed to provide.

A deterministic system inverts that. Verification isn't a task performed after the answer; it's built into the answer. Click any finding and you land on the document, page, and paragraph it came from. Trust doesn't have to be extended to the system. It only has to be extended to the source evidence — which is where a lawyer's trust belonged all along.

The standard the profession should demand

We think the bar for AI in legal evidence work is simple to state:

Every output cited to source, or no output at all. Confidence is not a substitute for provenance. A system that cannot show its working has no business in work that carries professional liability.

That's not a marketing position. It's how Pivot's architecture works — and in a profession where trust is earned slowly or not at all, we think it's the only foundation worth building on.