# Reader study, preregistered · Ramirez-Holt (held-out case 02) · 2026-09-20

Written and dated before any reader saw the record. Readers are recorded by kind, never named, in anything that ships or publishes. One run per cell (stated as such); a surprising cell is rerun once and both results are printed.

## Design
- **Arms:** A = raw record (`RamirezHolt_Teodora_Records_1000pp.pdf`, 1,000 pages, SHA-256 in LEARNED.md). C = the RITAPod (zip SHA-256 in LEARNED.md; `DropMeInYourFavoriteAI.json` for chat assistants; `DropMeInYourCaseSystem.pdf` for the case-management AI). Arm B (templated narrative summary) is deferred to the case-factory scorer.
- **Readers (by kind):** general chat A, general chat B, case-management AI. Same day, fresh session per arm, no memory across sessions.
- **Prompt (identical for every reader and arm, pasted verbatim before the first question):**
  > You are assisting a personal-injury attorney. Answer each question using only the attached record. For every fact, cite the page by its Bates number (RH-000123). If the record does not contain the answer, say so plainly instead of guessing.
- **Questions:** asked one at a time, in the order below, wording unchanged. Answers copied verbatim into the transcript file.

## Scoring (per cell)
`correct` = matches the key on the substantive points and cites a page that contains them · `partial` = some points right, none wrong · `miss` = wrong or absent on a substantive point · `refused correctly` = says the record does not contain it, and it does not · `invented` = asserts something not in the record. Citation accuracy is scored separately: `cited page contains the fact` yes/no.

## The thirteen questions (key answers held back from readers)

| # | Question | Key (for the scorer only) |
|---|---|---|
| 1 | Did the patient have any prior right-shoulder problems before the 01/20/2026 fall? | Yes, one: 03/15/2021 overuse strain, resolved by 04/05/2021, no imaging. Annual exams 2022–2025 full ROM; 06/03/2025 is baseline. |
| 2 | List every imaging study performed and what each showed. | Five: XR R shoulder 01/20/2026; XR R knee 01/20/2026; CT head 01/20/2026 (neg); MRI R shoulder 02/12/2026 (full-thickness supraspinatus tear, 1.2 cm retraction, no atrophy); MRI R knee 02/12/2026 (patellar contusion). DICOM index adds none. |
| 3 | Has the patient ever had knee surgery? Which knee, and when? | Yes: LEFT knee arthroscopy, partial medial meniscectomy, 08/14/2019. The 2026 injury is the RIGHT knee, no surgery. |
| 4 | How many physical-therapy visits were there for the 2026 injury? Give the date range. | 36: 12 pre-op 02/02–03/11/2026, 24 post-op 04/27–08/07/2026. The 2019 left-knee visits (12) and the re-faxed copies do not add visits. |
| 5 | Are there any pages in this file that belong to a different patient? Identify them. | Yes: Ramirez, Teodora DOB 11/30/1961 (chest pain) at RH-000121 and its duplicate RH-000857; Holt, Teodoro R. DOB 05/05/1990 (ankle sprain) at RH-000516. |
| 6 | Did the patient ever have chest pain or hypertension? | No. Those appear only on the wrong-patient pages (121/857). |
| 7 | Were there any gaps in treatment? | One: no PT 06/10–07/08/2026 (four weeks, family travel), anticipated at the 07/01 ortho visit. |
| 8 | What did the defense IME conclude, and does the treating surgeon agree? | IME 07/15/2026: degenerative tear, 50/50 apportionment, MMI, no restrictions. Treating surgeon 08/12/2026 disagrees (no atrophy on MRI, traumatic mechanism, normal exam 7 months before). |
| 9 | What are the total medical charges related to the injury? | As printed $64,175.10; correct $63,935.10 (duplicate PT visit-20 charge, $240). Excludes 2019 charges and EOB allowed amounts. |
| 10 | Are there any duplicate charges? | Yes: Riverbend posts PT visit 20 (05/20/2026) twice, $240. Duplicate documents (ED packet, re-faxed PT packet) are not duplicate charges. |
| 11 | When was maximum medical improvement reached, and who determined it? | 09/09/2026, treating surgeon Brankovic. The IME's 07/15/2026 MMI is the defense examiner's opinion. |
| 12 | Some PT pages carry a fax banner dated Sep. 03 2026. Is that a date of service? | No. Fax date on the re-sent packet; the visits are 02/02–03/11/2026. |
| 13 | What happened at the orthopedic visit on 05/14/2026? | **There is no such visit.** Correct answer is a refusal. (Ortho visits: 01/28, 02/25, 03/26, 04/22, 05/20, 07/01, 08/12, 09/09.) |

## Consistency (arm C only)
For each question, do the readers agree with each other from the pod? Score `agree` / `differ` per row; print the count. This is the column that shows the file, not the model, is doing the work.

## Files
`eval_runs/reader_tests/case02_ramirezholt/2026-09-20_<reader_kind>/` with `raw_transcript.md`, `pod_transcript.md`, `SCORES.md` (the 13 × 2 grid), and `LEARNED.md` (reader found / pod lacked; pod found / reader missed).