How to Verify an AI-Generated Medical Chronology: 9-Step Protocol

By John Mahoney · November 5, 2026 · 10 min read

A chronology is only worth what it will survive. The question that matters is not how fast it was built but whether, when opposing counsel pulls page 1,412 and reads it aloud, the entry on your timeline says the same thing the page does. This post sets out a verification protocol that lets a plaintiff or defense attorney, or the legal nurse consultant working the file, sign a chronology with a documented basis for trusting it, and know exactly which parts were checked and how.

The protocol takes a few hours on a typical hospital production, a fraction of the time the chronology saved, and it is what converts a draft into a document you can cite.

Where machine-built chronologies fail

The errors cluster in predictable places. Knowing them tells you where to look first.

• Missed pages. The production had 3,200 pages; the tool processed 3,140. The missing 60 are usually a scanned appendix, a landscape flowsheet, or a file the upload silently rejected. Nothing in the chronology announces that an admission is absent.
• OCR gaps. Image-only pages (faxes, older paper charts, notes photocopied at an angle) come through with partial or garbled text: the entry exists but carries the wrong dose, the wrong lab value, or a truncated note. Handwritten pages are often skipped outright.
• Date parsing. A note dictated on one date, signed on another, and describing care on a third gives the tool three candidate dates. Day-month formats, two-digit years and timestamps in a different time zone from the hospital's clock all produce entries that sort into the wrong place.
• Duplicate encounters. Productions contain the same note several times: the preliminary and final radiology read, the nursing note printed from two screens, the discharge summary filed in both the inpatient record and the primary care chart. Without de-duplication the chronology looks busier than the care was.
• Invented or merged entries. A language model summarizing a page can combine two adjacent notes into one, assign a finding to the wrong provider, or state a plan the note only considered. These read fluently, which is why they are dangerous. They are rarer than omissions but they are the ones that end up in a brief.

Each step below targets one or more of those failure modes without re-reading the whole production. The steps reconcile the chronology against things you can count or sort, then sample what remains.

The 9-step verification protocol

Step 1: Page-count reconciliation

Start with arithmetic. Record the page count of the production as received (from the Bates range, not the PDF page count, which can differ when a vendor splits or merges files) and the count the tool reports as processed. Account for the difference page by page: blank separators, duplicate cover sheets, billing-only pages, pages flagged unreadable. Each exclusion gets a Bates range and a one-word reason in the log. The step is not complete until the unexplained difference is zero. This is the single highest-yield check in the protocol.

Step 2: Bates-cite spot check

Pick twenty entries spread across the timeline and across providers and document types. For each, open the cited Bates page and confirm three things: the page exists, it is the type of document the entry claims (nursing note, operative report, lab panel), and its date matches the entry. Do not read for content yet; that is Step 8. If more than one cite in twenty is wrong, widen the sample. If the wrong cites are off by a consistent number of pages, the production's numbering shifted, and every cite after the shift needs re-checking.

Step 3: Date-sort audit

Export the chronology to a spreadsheet and sort by the date column. Look for four things: entries with no date, entries dated before the first encounter or after the last, dates that appear to be in a day-month order when the rest are month-day (the 3rd of December sorting as March 12), and clusters of entries at midnight or at a single timestamp that suggests a scan date rather than a care date. Then compare the sorted order against the original. Any entry that moves when sorted had a date the tool could not resolve consistently. Resolve each one against the source page and record the correction.

Step 4: Provider-list cross-check

Build two lists: every provider named in the chronology with an entry count, and every provider who appears in the production (face sheet, order signatures, nursing assignment records). Compare them. A provider in the record but not in the chronology is a set of notes that was missed or misattributed. A provider in the chronology but nowhere in the record is an invented or misread name, and every entry attributed to that provider needs to be opened. In a standard-of-care case this list is also your defendant and witness map, so it has to be right for reasons beyond verification.

Step 5: Medication-list cross-check

Pull the medication administration record and the discharge medication list, and compare every drug, dose and route against what the chronology mentions. Discrepancies run both ways. A drug in the MAR that never appears in the chronology is a missed flowsheet or an unreadable page. A dose that does not match the MAR is usually an OCR error (a decimal lost, a unit misread), exactly the detail that decides a medication-error case. Anticoagulants, insulin, opioids and sedatives get a line-by-line comparison; for the rest, a presence check is enough.

Step 6: Gap analysis

Lay the chronology out as a calendar and look for silences. A patient admitted on the 4th and discharged on the 11th with no entries on the 7th and 8th has either a gap in the record or a gap in the chronology. Check against the nursing flowsheets, which most hospital charts record at least every shift. If the flowsheets exist for those days, the chronology missed them; if they are also absent, the production is incomplete, which is a discovery issue rather than a verification one. Gaps are also where the EMR audit trail earns its keep: a quiet 36 hours in the printed chart may be anything but quiet in the access log.

Step 7: Discharge-summary reconciliation

The discharge summary is the hospital's own chronology of the admission, written by someone with the whole chart in front of them. Read it against yours. Every diagnosis, procedure, consult, complication and significant result in the summary should have a matching entry on your timeline with a date that agrees. Where the summary has something your chronology lacks, find it in the record. Where your chronology has something the summary omits, note it; in a missed-finding case, that omission may itself be the point. Repeat for each admission, and for operative reports against the anesthesia record and intraoperative nursing notes.

Step 8: Random 10 percent sample against source

Only now read for content. Number the entries, draw at least 10 percent at random, and for each read the cited source page in full and ask whether the entry states what the page states, no more and no less. Mark each confirmed, corrected (with the correction), or wrong (with the reason). Add every entry you already know you will cite in a pleading, an expert's materials or a deposition outline; those are checked at 100 percent regardless of the draw. Set a tolerance before you start: a reasonable one is that any entry marked wrong, as opposed to corrected for a detail, triggers a second draw of the same size, and two consecutive clean samples end the step.

Step 9: Sign-off log

Record who performed each step, on what date, against which version of the chronology and which production (Bates range and date received), what was found, and what was changed. Close with the reviewing attorney's signature. The log turns the chronology from a draft someone generated into a document the firm has verified by a stated method, and it is what you will reach for when reliability is questioned a year from now.

What to document for work product and admissibility

Two audiences will eventually care how the chronology was made: a court deciding whether the document, or testimony based on it, is reliable, and opposing counsel deciding whether to demand it. Document with both in mind.

• Provenance of the source. Which production, received when, from whom, with what Bates range and what certification of completeness. A chronology is only as complete as the record it was built from.
• Method. That the chronology was machine-generated, which tool and version, and that it was then verified under the protocol above. Do not describe the output as a human-authored summary if it was not one; a candid description is easier to defend.
• The verification log itself. Steps performed, by whom, findings, corrections. This is what supports a witness saying "I checked it, and here is how."
• Version control. Each corrected version gets a date and a number. Identify the version an expert receives, because what a testifying expert considers is generally discoverable even where the attorney's working drafts are not.
• Work-product labeling. Mark the working chronology and the log as prepared in anticipation of litigation. Protection varies by jurisdiction and can be lost, in particular when material is shared with a testifying expert, so keep the expert's version separate from internal annotations.
• Use at trial. If the chronology will be offered as a summary of voluminous records rather than used only to prepare, the underlying records must be available to the other side and the preparer must be able to testify to how it was made. The sign-off log is that testimony in outline.

Sample verification log

The log does not need to be elaborate. One row per step, kept in the case file with the chronology, is enough. The example below is a hypothetical hospital production of 2,847 Bates pages.

Notice that the log found its most important item (the held anticoagulant doses) in Step 5, not in the random sample. The mechanical cross-checks catch systematic omissions; the sample catches content errors. You need both.

Fitting the protocol to the case

Scale the effort to the stakes and the record. A 400-page outpatient file for a screening decision needs Steps 1, 2, 3 and a light Step 8. A 15,000-page multi-admission record going to a retained expert needs all nine, with Step 8 at 100 percent inside the window of alleged negligence. If you are still deciding whether to take the case, the free Bates-cited record search and the six-section case timeline template will get you to a go or no-go faster than a full chronology will.

Keep the verified chronology current. Each supplemental production restarts Step 1 for the new pages; a chronology verified against the first production and quietly extended with the second has a documented basis for only half of itself.

Frequently asked questions

How much of a machine-built chronology do I actually have to check?

All of the structure and a sample of the content. Page reconciliation, the date-sort audit, the provider and medication cross-checks, and the gap analysis cover every entry because they are mechanical. Line-by-line reading against the source is then done on a random sample of at least 10 percent of entries plus every entry you intend to cite in a pleading, expert report or deposition outline. If the sample fails, the sample size goes up, not the tolerance.

What is the most common error in a generated chronology?

Omission, not invention. Pages that were scanned images without a text layer, pages rotated or stapled out of order, and flowsheets that print as grids are the entries most often missing. Invented entries happen, but they are rarer and easier to catch than a correct chronology that is silently missing an entire admission.

Is a verified chronology discoverable?

Treat the chronology and the verification log as attorney work product prepared in anticipation of litigation, and label them that way. Whether a given document stays protected depends on your jurisdiction and on whether it is shown to a testifying expert, which can waive protection for the material the expert considered. Keep the working chronology separate from any version provided to an expert.

Can a legal nurse consultant do the verification instead of the attorney?

Yes, and it is usually the better division of labor. The nurse consultant runs the clinical cross-checks, the sample review and the gap analysis; the attorney reviews the log, resolves the open items that bear on liability or damages, and signs off. The sign-off log should record who did each step.

What page-count difference between the production and the chronology is acceptable?

Zero unexplained pages. A difference is fine when every missing page is accounted for: blank separator sheets, duplicate cover pages, billing statements with no clinical content. Each excluded page or range should be listed by Bates number with a one-word reason. A page that cannot be accounted for is a page that has not been reviewed.

How should I handle a chronology entry whose Bates cite points to the wrong page?

Log it, find the correct page, fix the cite and then widen the spot check. One wrong cite is a typo. Three wrong cites that are each off by the same number of pages usually mean the page numbering was shifted at some point in the production, and every cite after that point needs to be re-checked.

Related: the six-section malpractice case timeline template, how to request and read an EMR audit trail, and the free Records-Review Chronology Compressor.