Medical and legal document synthesis

A synthesis run on a medical or litigation file is finished when every sentence in the briefing either points at a page in the record or is left blank. The model does not reserve the claim, authorize a payment, diagnose an injury, or find liability. You still do those things, on the strength of the file and the coverage, after you have read the cites.

Complex bodily injury and litigated claims arrive as stacks: hospital charts, IME and peer-review reports, bills, demand letters, pleadings, discovery, expert reports, and correspondence. The briefing puts the facts you will use in one place, with a page you can open. It is not a second opinion and it is not a settlement recommendation.

Treat any sentence without a page cite as unfinished work. If the draft says the claimant has a herniated disc at L4-L5 and there is no Bates or chart page on that sentence, delete the line or attach the page.

You will usually start this work after AI FNOL intake and claim triage has assigned the file to a complex or litigated queue. Intake can route. It cannot read a medical chronology or a production set for you.

Load the records you intend to cite, and only those records. Pull the medical chart, the litigation file, the expert reports, and the correspondence from the system you already use. Policy admin and claims suites such as Guidewire and Duck Creek hold the claim number, coverage, and diary. Document-review platforms such as Relativity and Everlaw hold productions, depositions, and work product. The model reads what you load from those systems. It does not replace them.

Keep the load bounded. A briefing that mixes the wrong claim, an unsigned draft report, or a prior-policy medical file will cite pages that do not belong to this loss. If a document is not in the load, the briefing cannot cite it. If you need a missing record, request it. Do not let the model fill the gap.

Ask for a draft in a fixed shape: loss date, providers, procedures and imaging, work status, billed charges, and pleadings or expert opinions, each as stated on a named page, with blanks for open questions. Every factual sentence carries a cite your desk already uses: Bates, PDF page, chart date plus page, or production ID.

A usable line names the page: "On 14 March 2024, MRI lumbar spine at Riverside Imaging reports disc herniation at L4-L5 contacting the left L5 nerve root (Riverside Imaging, 14 Mar 2024, p. 3)." You can open page 3 and see the same words. A sentence that says the MRI confirms a surgical disc, with no page, is an unsourced gloss. Drop it.

Walk one file. On a litigated auto bodily-injury claim you load the discharge summary, MRI report, plaintiff's expert report, answer, and last three demand letters, not a prior claim's chart. The draft returns lumbar strain cited to discharge page 2, MRI herniation cited to imaging page 3, relatedness of surgery cited to expert report page 8, and a causation denial cited to answer page 4. MMI date stays blank; no page states it. You do not adopt the expert's relatedness opinion as the claim's finding. You set the reserve yourself after opening the cites, and you do not pay from the draft.

Spot-check cites before you keep the draft: open the page, confirm a fair reading, and confirm the page is from this claim. If a cite is wrong, strike the sentence.

Computer vision damage assessment may already sit on the same file for vehicle or property photos. Keep those outputs in their own lane. Photo estimates do not diagnose a spine.

Empty stays empty. If the load has no page that states a diagnosis, a date of maximum medical improvement, a wage figure, a prior accident, or a legal finding, that briefing field stays blank. A blank is a request for a record, or a decision you will make without that fact. A filled blank is an invention.

Do not let the model interpolate visits that are not in the chart. Do not let it infer a diagnosis from a CPT code, a drug name, or a therapy plan unless a named page states that diagnosis. Codes and medications are facts when they appear on a bill or a list. They are not a clinical impression.

The same rule applies to the legal file. If no pleading, order, or expert report in the load states a finding on liability, causation, or damages, the briefing does not state one. "Likely comparative negligence" with no page is advocacy the model is not entitled to write.

When a page is ambiguous, quote the ambiguity and cite the page. "The 12 June clinic note lists low back pain and does not name a disc level (Clinic note, 12 Jun 2024, p. 1)" is a briefing sentence. "The claimant has a lumbar disc injury" is not, if that is all the note contains.

A briefing that looks complete because every field has text is worse than a briefing with holes you can see.

The model does not invent a diagnosis. It may copy a diagnosis that a named provider wrote on a named page, in that provider's words or a close paraphrase, with the cite. If two providers disagree, the briefing shows both cites. It does not pick a winner.

The model does not invent a legal finding. It may copy what a complaint alleges, what an answer denies, what an order recites, or what an expert report concludes, each with a page. Your coverage analysis, your liability view, and your evaluation remain yours.

Quality means a briefing sentence that cites the record page, and a blank when the page is silent, not a recommended reserve or a payment instruction.

A fluent medical summary is not the treating opinion, and a fluent liability paragraph is not the file's legal conclusion. Strike confident language with no page, or a page that does not say what the sentence claims. If you need a medical opinion the file does not contain, request an IME or a records review. If you need a legal conclusion, ask counsel or write it after you have read the cites.

You still set the reserve and you still pay. The briefing is an index into the file, not a trigger for straight-through claims payment. Auto-pay from a synthesis draft is out of scope. A well-cited chronology can still omit a coverage issue, an exclusion, a prior settlement, or a fraud referral that lives in another system.

If cited lines show a date of loss that does not match FNOL, treatment that predates the accident, or billing that does not match the operative report, send them to the desk as questions. Pattern work belongs with real-time fraud pattern detection, not with a prompt asked to explain the discrepancy away.

Do not paste the briefing into the reserve field. A reserve is your estimate of cost, based on coverage, facts, and judgment. Treating the brief as the reserve fails even when every sentence is cited, because the brief does not know limits, deductibles, or your view of liability.

You load, you generate, you spot-check cites, you delete unsourced sentences, you leave blanks, and you decide. If a later record arrives, load it and revise the lines that record supports. For each sentence, open the cited page. If the page says it, keep the sentence. If the page is silent, the field stays empty. Then you reserve, you negotiate, or you pay, using the file, not the model's confidence.