Tested on synthetic records only. No real client record has been through this tool. Every patient, provider and page shown here is invented.

Medical chronology Synthetic records only

Scanned records in. Every entry checked against the page it cites.

Today the chronology is run as a reviewed service: a firm sends records, and a reviewer approves every entry beside its page. OCR runs on the machine that holds the records. To draft entries, the text of each visit's pages is sent to an outside model service; for client records that step stays off until a business associate agreement is signed. So far only synthetic records have been through it.

To decide

Cedar Hollow Medical Center Emergency Department

Findings, word for word from the pages

  • Neck pain after a motor vehicle collision.
  • Tenderness over the cervical paraspinal muscles.
  • Cervical strain (S16.1XXA).
  • Ibuprofen 600 mg, follow up with primary care.

MOCK000001MOCK000002

Proposed by the model. Every word found on the cited page.

The pipeline

Eight stages. A model has a hand in one of them.

Each stage is done by code, by a model or by the reviewer. The model's only job is to propose. It decides nothing.

  1. Scan in

    The firm's PDFs are received once an authorization is on file. Every page gets a Bates number and a row in the page list.

    Code
  2. Read locally

    OCR runs locally, on the machine that holds the records. Blurred, faint and handwritten pages are marked low.

    Code
  3. Propose

    A model drafts one entry per visit, quoting whole sentences from the pages. This is the one step where page text leaves that machine.

    Model
  4. Check each word

    Every proposed word is compared with the text of the page the entry cites.

    Code
  5. Review beside the page

    The reviewer approves, rejects or edits each entry with the page image next to it.

    Reviewer
  6. Deliver

    Only when every entry has a decision and every page is accounted for. Word, PDF and Excel.

    Code
  7. Read back

    Each file is opened again and compared with the entries the reviewer approved.

    Code
  8. Purge

    The records, the entries and the service's copy of the report are deleted, 30 days after delivery unless another time is set.

    Code

Watch how it works

The whole tool in seventeen short scenes.

An animated walk from intake to purge, including four kinds of entry the checker refuses. Use the arrows to step through at your own pace. On a narrow screen the same steps are shown as a list.

Synthetic records only. The explainer follows a different invented case from the demo.Open the explainer on its own page

What the checker refuses

An entry must be made of the page's own words.

Findings are quoted, not paraphrased, so code can check them. An entry is refused if any of these is true. A refused entry goes to the reviewer's list with the reason. The reviewer fixes it against the page, and the fix is checked again.

A wrong date

The entry's date is not the date of service printed on its page.

Another provider's page

The entry mixes in a page from a different provider, even one seen the same day.

A cut finding

A quoted sentence stops short of what the page says.

An added word

A No in front of a finding, or any word the page does not have.

Another visit's page

The entry cites a page that belongs to a visit on a different date.

A page from another matter

Citations stay inside the matter they came from.

A visit type the provider contradicts

An entry called an emergency visit on a page from a physical therapy clinic.

An invented cost, or a dropped diagnosis

A billed amount the pages do not show, or an entry that leaves out its page's assessment.

Review a chronology
The review screen showing an entry marked Checker rejected, with the reason that it cites a page of a visit on another date, beside the scanned page.Open full size
A planted slip, refused: the entry cites a page from a visit on another date. Synthetic patient. The program's working name is Transcript Check, and the captures on this site are from code later than the preview build.
Review a chronology
The review screen listing seven pages flagged as a possible missed visit, with the message that the page reads as a visit and no entry cites it.Open full size
Pages that read as a visit with no entry citing them are flagged. Synthetic patient.

What it flags for the reviewer

Every page is accounted for before anything is delivered.

Visits the model missed

A page that reads as a visit, with no entry citing it, is listed as a possible missed visit. The reviewer writes the entry or says why none is needed.

Another patient's page

When most pages name one person and a page names someone else, it is flagged as possibly misfiled.

Duplicate pages

A second copy of a page is suggested as a duplicate. It keeps its Bates number and stays in the page list.

Pages OCR could not read well

No entry may cite a low page until a person has checked its text. For a handwritten page, the reviewer types what it says and the report marks the text as typed by the reviewer.

Bills and totals

Bill pages are matched to their visit. When the reviewer adds a bill, its amount reaches the report, and one bill cannot be counted twice.

An encrypted disk, or nothing opens

The review screen refuses to open records on a disk that is not encrypted.

Review a chronology
The review screen showing a page that names another patient, flagged as possibly misfiled, with buttons to write an entry, mark no clinical content or mark a duplicate.Open full size
A page naming someone else, flagged in the demo records. Every name is invented.
Review a chronology
The review screen with eleven entries to decide and the first entry open beside its scanned page.Open full size
Eleven entries for eleven visits in the demo records, each waiting for a decision.

Delivery and purge

Delivered files are read back. Then the records go.

Delivery waits until every entry is approved or rejected and every page has an entry or a reason. The chronology is written as a Word document, a PDF and an Excel sheet from approved entries only.

Each file is then opened again by code and compared with what the reviewer approved. If a file does not match, it is not offered.

After delivery, the records, the entries and the service's copy of the report are deleted on a set date, 30 days later by default. The page list and the reviewer's decisions about each page are kept as a record of what was done.

Review a chronology
The delivered screen: each file was read back and matches the approved entries, with DOCX, PDF and XLSX buttons, a purge date and a Purge now button.Open full size
The demo matter after delivery, with its purge date. Synthetic records.

Measured, in our own tests, on synthetic records

Results from 42 synthetic cases.

2,824 pages and 1,136 visits. On 10 Nov 2026 we re-ran the checks, delivery and purge on them (OCR text for 37 of the 42 cases was kept from the 7 Nov run). This run used no model: the entries tested were exact quotes from an answer key, plus errors planted on purpose. It measures the checker, not a model's drafting.

Planted errors the checker refused
5,368 of 5,368
Eight kinds of error. 10 Nov 2026. 3
True entries that passed the checker
1,128 of 1,136
8 true entries were refused. All trace to OCR reading a bracket as a letter. 10 Nov 2026. 3
Bill pages found and matched to their visit
132 of 132
None matched falsely. 10 Nov 2026. 3
Lost visits flagged as possibly missed
15 of 15
10 Nov 2026. 3
Duplicate pages found
48 of 48
10 Nov 2026. 3
Cases delivered, read back and purged clean
42 of 42
10 Nov 2026. 3

Limits

What the chronology cannot do yet.

This is the least proven of the three tools. Read these before anything else on this page.

  • Every record so far is synthetic. No real client record has been through the tool. The sorting rules fit our own synthetic layouts.
  • The large test used no model. It tested the checker with perfect entries and planted errors. With a model, we have one small run: 28 synthetic pages gave 11 entries for 11 true visits on 10 Nov 2026.
  • 8 true entries were refused. OCR read ] as l in a medication list, so a true quote no longer matched its page. Refusing is the safe outcome, and it leaves the entry for the reviewer to fix.
  • Real handwritten pages cannot be cited by the model. On 22 pages of real, public 1963 hospital records, all 10 handwritten pages were marked low and OCR could not read them. A person has to type the text.
  • Real forms confuse the provider guess. On those same pages the suggested provider was wrong on 8 of the 13 pages that name one. It is a suggestion, and the reviewer corrects it.
  • Dates and document breaks slip on real pages. On the 9 typed real pages that carry a date, the date was read right on 7. The 22 real pages were split into 20 documents where 13 was right.
  • A slightly blurred page can still be misread. Heavier blur is marked low. Light blur may not be. The reviewer with the page beside the entry is the last check.
  • A matter cannot be started from a screen. Creating the matter and reading the files in is still done by script.
  • A misfiled page has no label of its own. A page from another patient is flagged, but the choices for it are still No clinical content, Duplicate or Write an entry.

See entries beside their pages.

The demo shows the invented plaintiff's 28 pages of records, with the checker's verdicts on true entries and on planted slips.