Built around the physician’s signature
Most record-review tools are built for the claims desk or the law firm. Lexamed is built for the evaluator who signs the report, and every design choice follows from that.
Who is a record-review tool built for?
A medical-legal report is a regulated document. The evaluator signs it, attests to the time spent and the pages reviewed, and answers for its opinions. A tool built for the claims desk produces a summary sized to a claim decision. A tool built for the evaluator has to produce something you can put your name on.
| Dimension | Claims-desk tools | Outsourced review | Lexamed |
|---|---|---|---|
| Built for | Adjusters deciding a claim | A physician who sends the reading out | The physician who signs the report |
| What comes back | A summary sized to a claim decision | A chronology prepared by someone else | A draft of your report, with your sections left blank |
| Who read the pages | The tool | Another reviewer | You, with every entry one click from its page |
Where does the model stop?
Four determinations make a medical-legal report yours. Lexamed is written to refuse all four. It collects the evidence each one rests on, cites it, and leaves the section blank. A draft that offered an opinion would be a draft you had to argue with instead of one you could finish.
Causation
Whether the injury arose out of and in the course of employment is never stated by the model.
Apportionment
What share of disability is industrial is left blank for you, with the prior history surfaced beside it.
Impairment
No whole person impairment figure is drafted. The findings that support one are collected and cited.
Maximal medical improvement
Whether the applicant is permanent and stationary is your determination, not a field the draft fills in.
Can every claim be checked?
Every chronology entry carries the page it came from and a verbatim quote from that page. Before any model is asked whether the entry is supported, the quote is matched against the page text. If it is not there, the entry stays flagged. It is never quietly dropped, and it is never quietly accepted.
The same rule runs through the record itself. A page served twice is grouped with its copy and counted, not hidden, and it is read once. A page with no date stays undated rather than being guessed into the timeline. A note that restates an earlier study is folded into the study it restates, so one MRI is one line.
What does it hand back?
Not a timeline to copy from. A draft in the shape the parties expect, with the history assembled and the evaluator’s sections waiting.
- The issues presented, transcribed from each party’s cover letter in the order asked.
- Every account of how the injury occurred, quoted, with the places the tellings drift from one another.
- Activities of daily living placed in the AMA Guides categories a rating rests on.
- A records-reviewed accounting the draft checks against the record set before you sign the § 4628 declaration.
- The Medical-Legal Fee Schedule service items the review supports, at the published amounts.
How do you know it works?
Quality is measured, not asserted. Each of these exists in the product today and runs before a change reaches you. The numbers they produce can be reproduced, and the limits of those numbers are written down beside them.
A thresholded benchmark
Page counts, classification, duplicate grouping, encounter recall, citation integrity and grounding, each with a pass line a release is reviewed against.
Planted-issue cases
A physician writes down what is wrong with a synthetic record set first. The benchmark then checks whether the pipeline caught each planted issue.
A report judge
The finished draft is checked as a document: that it agrees with itself, with the record set, and with the approved wording for empty sections.
A written launch gate
Any model or prompt change is run through the benchmark, replay against reviewer decisions, blind side-by-side preference and a workflow suite. A person reads the results and decides whether it ships.
What happens to your work?
Nothing you upload or correct trains a model. Your corrections tell us which synthetic test cases to build next, and that is all they are used for. An entry you have not touched is never counted as accepted. The log of who did what is append-only and outlives the case it describes. A change to how the model is prompted for your practice needs a person’s approval and a limited rollout before it applies to every case.
What does it not do yet?
Written here so that nobody has to find out later.
- Quality numbers come from planted cases
- Every benchmark figure so far comes from physician-authored synthetic record sets. The citation check is there so you can judge each result yourself.
- Real patient records are not accepted yet
- The free trial runs on synthetic records only. Real-record access opens after the business associate agreement and the hosting move are complete, and the product enforces that, not just the terms.
- Scale is unproven
- The acceptance target is a three-thousand-page production in under thirty minutes. The largest case measured so far is a fraction of that.
- Degraded scans are hard
- Faxed, rotated, low-resolution pages are read with real optical character recognition and flagged when confidence is low. Recovery on the worst pages is partial.
- No practice-system integrations
- Records arrive as PDF uploads. There is no document management or case management connection yet.
Questions physicians ask
- Who is Lexamed for?
- Physicians who write medical-legal reports, such as Qualified Medical Evaluators, Agreed Medical Evaluators and independent medical examiners, and the staff who prepare records for them. The product assembles the history so the physician can spend the review forming opinions.
- Does the AI write any part of the medical-legal opinion?
- No. Causation, apportionment, impairment and maximal medical improvement are never drafted by the model. Those sections are left for the evaluator, with the supporting evidence collected and cited beside them.
- How is a citation checked?
- Each chronology entry carries the page it came from and a verbatim quote. The quote is matched against that page’s text before any model judgement, and an entry whose quote is not on its page stays flagged rather than hidden.
- Is my data used to train the model?
- No. Reviewer corrections inform which synthetic test cases get built next. They are never used as training data, and the audit log of who did what is append-only.
- What does the free trial include?
- The full review workspace on synthetic productions, with nothing to set up and no card at signup. Commercial pricing is published before the trial converts to a paid plan.