AI · Language models in device matters
A method for large document sets.
A device matter arrives as tens of thousands of pages: productions, depositions, complaint files, device logs, FDA data. We use large language models to index, extract, and cross-reference that material, under written instructions that make every statement checkable. The instructions are published here.
The method
Fast reading, with every line cited.
A language model reads a production in hours, not weeks. It also makes mistakes with confidence. The method keeps the first and removes the second: the model works from an indexed copy of the documents, writes nothing without a document and page, and a second pass opens every citation before anything leaves the team.
Four rules hold throughout. The originals are never touched; the work is done on an extracted, indexed copy. Every statement carries a citation to a document and a page, or a deposition page and line. An empty field means the record does not say it; a guess is never written in its place. And the output states what the records say and where, not what caused the event or who is liable. Those opinions belong to counsel and to the retained expert.
Before any case material goes into a model, settle the confidentiality questions: the protective order, the data-handling terms of the model service, and where the extracted copy lives. Our guide 06, Case documents and language models, walks through them.
A model is a reader, not a witness. Nothing it produces goes to a court or to the other side until a person has opened the cited pages. Every skill below ends with that check.
Claude skills
Instructions Claude follows on a case folder.
A skill is a short set of written instructions that Claude follows when it works on a case folder: how to index the production, what to extract, how to cite, what to verify before it writes. These six are the methods we use ourselves, written so that every statement in the output can be opened and checked.
Each skill is a folder with one file, SKILL.md. Download the .skill file and add it in Claude’s skills settings, or unzip it into your skills folder. The skills work together: the case brief builds the archive the others read, the deposition summaries feed the comparison, and the timeline and the complaint reader take the FDA data from the search.
- 01
Case brief case-brief
One brief from the whole production: what happened, what each witness said, what documents exist and what is missing, the legal Q&A, the engineering record, and two timelines, every line cited to a page.
Output Brief (Markdown) + index, timelines, witnesses, legal Q&A as CSV
- 02
Deposition summary deposition-summary
One transcript into a high-yield summary by topic, with page and line on every statement: admissions, denials, the documents shown, conflicts with the record, what the witness could not say, and what was never asked. Beside it, a searchable testimony database, one row per answer.
Output Summary (Markdown) + testimony database, exhibits, topics (CSV)
- 03
Deposition compare deposition-compare
Several witnesses side by side, fact by fact: who affirmed, denied, qualified, did not know, or was not asked, with the documents in a column of their own. Differences in the witnesses’ own words, vantage points, who knew what when, and the questions for the next witness.
Output Comparison (Markdown) + fact matrix, differences, next questions (CSV)
- 04
Device timeline device-timeline
Four streams on one time axis: the device, the field, this unit, and the case. Each entry has a document date and a knowledge date, and the MDR, 806, and CAPA clocks are checked against the records.
Output Timeline (Markdown + CSV), clocks table, generated figure
- 05
Complaint reader complaint-reader
Complaints, MDRs, service records, and CAPAs as one table with the same fields. Series found across years and models, each complaint reconciled with its MDR, each CAPA linked to the complaints before and after it.
Output Record table, series, MDR reconciliation, CAPA links (CSV) + summary
- 06
openFDA search openfda-search
FDA’s public device data through the API, pulled into files: adverse event reports, recalls and enforcement reports, 510(k) and PMA records, classification, and UDI. Counts by year, type, model, and problem, raw and deduplicated, each with its query and date.
Output Search record, data (JSONL + CSV), counts, summary
In practice
How a matter runs with them.
- 01
Index
The production, the depositions, and the FDA data become one archive: one text file per document, page markers, an index with a category per document. The case brief skill builds it; the openFDA search skill adds the public record.
- 02
Extract
Timelines, witness testimony by topic, the fact matrix across witnesses, the legal Q&A, the complaint series: tables first, each cell cited, then prose written from the tables.
- 03
Check
Every citation opened and read before the output is used: the page, the line, the words. What does not check out is removed, not softened.
- 04
Engineer
The brief tells our engineers where to look. The device, the logs, and the bench tests tell them what happened. The two are kept apart in the report, and both are cited.
Get in touch
Tell us about the device.
Share a brief overview of the device, the question you need answered, and any deadlines. We’ll explain how we can help and recommend the next steps.
info@alphadevices.io