Patenting the forensic chronology

In June we filed our seventh patent application. Seven inventions, each filed in both the United Kingdom and the United States. When I mention this, the assumption is usually that we have patented some clever model, a fine-tune, a prompt technique. We have not. None of the seven covers a model. All of them cover the machinery around the model, and that is the point of this post.

What a model cannot be made to promise

Large language models are extraordinary at reading, organising and describing. What they cannot be made to do, by any amount of prompting, is guarantee. Did you read that faxed page correctly, given that the scanner mangled half of it? Did anything enter this chronology that should not be evidence? Did the drafted report cover every record it was supposed to cover, and every question it was instructed to answer? A model will give you confident answers to all of these. It cannot give you guarantees, because guarantees are not in the gift of a probabilistic system.

In most domains that is acceptable. In medico-legal work it is disqualifying. An expert report is signed under a statement of truth. A chronology is only useful if every entry can be traced to its source. The questions that matter are exactly the ones a model cannot promise about its own behaviour. So the engineering that matters is the machinery that converts model behaviour into enforced properties: validators, structural invariants, audit records, architectural separations. That is what we filed on.

Seven inventions, one per layer

The filings follow the pipeline from a scanned page to a signed report. They cover how the system reads material the scanner has damaged; what it allows into a chronology, and what it visibly excludes; how its analysis is prevented from quietly changing its mind; how a drafted report demonstrates its own completeness, and its coverage of the instructions it was given; what the in-case assistant is permitted to reason over; and how accuracy is measured without patient data ever leaving the protected environment.

The detail belongs in the applications, not this post. At every layer there is a question a court could ask that a model cannot answer about itself, and at every layer we built the machinery that answers it. That machinery is what we filed.

Where the industry is heading

Most document-AI products today are thin layers over the same handful of foundation models. As those models improve, whatever advantage lived in the layer evaporates: prompting techniques circulate within weeks, fine-tunes depreciate with every base-model release, and the demos converge. Model capability is no longer where differentiation lives, because model capability is something every company rents from the same suppliers.

The differentiation that survives sits in system properties the model cannot supply: whether output is verifiable, whether completeness can be demonstrated, whether provenance survives from source to conclusion. That machinery is expensive to build, and it only pays for itself where the output is evidence, where a chronology entry may face cross-examination years after it was produced. A horizontal product serving every industry cannot justify the engineering. A vertical platform built for medico-legal work, operating under CPR Part 35 and UK data-protection law, cannot function without it. That asymmetry is why we expect this ground to stay ours.

What makes a chronology forensic

We use a specific term for what the platform produces: a forensic chronology. A chronology is a list of clinical events in date order; any careful reader of records can produce one, and by now most AI systems can produce something that resembles one. A forensic chronology carries a further property: every entry is prepared on the assumption that it will be challenged. Its date, its wording and its source must each survive the question “how do you know?”, asked years later, by someone whose job is to find the weakness.

That standard decomposes into separate engineering questions. Did every in-scope record contribute? Is each event dated by the event, rather than by a date mentioned near it? Does every entry trace to a source passage that still supports it? ClinLexis engineers each of these as its own guarantee, enforced in code where enforcement is possible and measured against expert review where it is not. That is the sense in which we use the word forensic: each entry is checkable, and checked.

The result is a chronology of exceptional forensic accuracy. That claim rests on architecture and measurement rather than on a model. Both are verified daily in live expert use, and both are inspectable by any customer who wants to test them.

The specification courts are writing

The most interesting thing about this portfolio is how closely it tracks what courts are starting to require. The Irish High Court’s new Practice Direction demands that AI output used in court documents be independently verified under human supervision, and singles out chronologies and expert reports. The Civil Justice Council in England and Wales is moving towards proportionate transparency for expert evidence. Verification, provenance, completeness, disclosure. Courts are increasingly requiring what we already engineered for.

More on the portfolio is on our company news page.