source/processed_csv/SUMMARY.md

Medical Exam CSVs — Build Summary

Built from data/updates_raw/ → data/processed_csv/. Nothing under updates_raw/ was modified; nothing was committed.

Totals: 58 CSVs, 7,812 questions, 179 extracted images, 168 image references — all resolve to real files.

CSV columns (all files): exam_id, question_number, question_id, question_text, choice_a, choice_b, choice_c, choice_d, choice_e, correct_answer, image_file, language, country, year

Per-exam

Exam	CSV	Questions	Answer-key source	Images
SSM Italy 2025	ssm_italy_2025.csv	140 ✓	PDF header: correct answer is always A	13
SSM Italy 2026	ssm_italy_2026.csv	140 ✓	PDF header: correct answer is always A	9
MIR Spain 2025	mir_spain_2025.csv	210 ✓	Official FSE JSON key (plantilla_definitiva_OFICIAL_v0.json)	25
MIR Spain 2026	mir_spain_2026.csv	210 ✓	Official plantilla_definitiva.pdf (v0 booklet canonical)	25
EDN France 2023	edn_france_2023.csv	305 (279 transversal + 26 LCA) ✓	None exists — see below	29
EDN France 2024	edn_france_2024.csv	339 (310 + 29 LCA) ✓	None exists — see below	40
EDN France 2025	edn_france_2025.csv	318 (289 + 29 LCA) ✓	None exists — see below	34
PNA Portugal 2025	pna_portugal_2025.csv	150 ✓	Official chave.pdf ("Prova A" column), 150/150	3
PES Poland (50 exams)	pes_poland_<specialty>_<year>_<wiosna|jesien>.csv	120 each = 6,000 ✓	Each exam's (klucz) key PDF, 6,000/6,000	1

Per-paper EDN counts match the source manifests exactly (93/93/93, 105/107/98, 93/98/98). The 878 "expected" EDN questions are the transversal papers; the extra 84 rows are the LCA sections.

Answer keys

• SSM: correct_answer = A for all 280 rows, per the header printed in both PDFs and consistent with the ☒/X marks on option A in 140/140 questions per year.
• MIR: annulled questions 13, 50, 64, 139, 142, 161, 208 are marked ANNULLED in both years — verified identical in the 2025 official JSON (nulls) and the 2026 official key PDF (blank cells). MIR exams have 4 options, so choice_e is empty. All 420 answers cross-checked against the keys.
• EDN: correct_answer is empty for all 962 rows. The CNG publishes questions-only PDFs with no official corrigé; community corrections exist only inside interactive prep platforms (edn.chat, oreille-externe.fr, refoua.fr) that require a live browser session. No downloadable key exists.
• PNA: 150/150 answers from chave.pdf.
• PES: 6,000/6,000 answers. 11 questions whose klucz marks them annulled keep correct_answer = X (list under Issues).

Images

Extracted with PyMuPDF, saved as PNGs under images/<exam_id>/, linked by bare filename in image_file (two files for one question are ;-separated).

• SSM: 22 figures, mapped geometrically (image position vs. question header on the page); spot-verified visually (2025 Q33 chest X-ray, Q43 lichen planus skin photo).
• MIR: 25 per year from imagenes.pdf, matched 1:1 to the image-linked questions (Qn → imagen_NN.png). Note the source defect fixed below.
• EDN: 103 figures; 92 questions linked, including all 24 QZP click-on-image questions.
• PNA: Q73 12-lead ECG; Q80 CTG + tocography traces (2 files).
• PES: essentially text-only — exactly 1 genuine figure (sleeve-gastrectomy diagram, chirurgia ogólna 2019 wiosna Q100).

Issues & uncertainties

1. PES 2021–2022 are scans, not text PDFs (contrary to the original staging notes). All 40 PDFs in fallback_nil_2021_2022/ are NAPS2 scans with zero text characters. Those 20 exams were OCR'd with tesseract (pol); their question/choice text carries OCR noise and is lower-fidelity than the 2018–2020 text-layer rows. Their keys were transcribed by hand from the key-page images (automated key OCR proved unreliable — systematic misreads), stored in ~/workspace/pes_parse/manual_keys/*.json, each validated to 120 entries.
2. PES scanned exams — empty choices: ~47 questions across 17 scanned exams have at least one empty choice_* cell where OCR couldn't split a choice cleanly. Correct answers are present for all. Worst: choroby_wewnetrzne 2021 jesień, choroby_wewnetrzne 2022 jesień, medycyna_rodzinna 2021 jesień, psychiatria 2022 jesień (4 each).
3. PES annulled (X in klucz): chirurgia 2021 jesień q114; chirurgia 2021 wiosna q15, q78; chirurgia 2022 wiosna q116; choroby wewnętrzne 2021 wiosna q93; medycyna rodzinna 2021 jesień q110; medycyna rodzinna 2022 jesień q105; medycyna rodzinna 2022 wiosna q51; pediatria 2021 wiosna q50; pediatria 2022 jesień q73, q79.
4. MIR 2026 image defect (fixed): the embedded ECG pixmap is mislabeled "Imagen 12" inside the image, but the booklet overlays the correct caption "Imagen 11". imagen_11.png is a 150-dpi render of the booklet page as candidates actually saw it; the other 24 are raw embedded images.
5. EDN question types: 81 QRP questions have 6–16 proposals — only A–E kept (extras dropped, listed in edn_france_build_issues.log). 119 QROC (open short-answer) and 24 QZP questions have no text choices by design. 26 TCS questions use the fixed 5-point Likert scale as choices A–E. 39 questions mention imaging/ECG only descriptively and have no figure — correctly left unlinked.
6. Minor source defects preserved as-is: MIR 2026 Q24 "variantepatogénicaenJAK2" (missing spaces in the PDF itself); PNA Q80 stem contains figure-axis text pulled from inside the CTG image; one dropped apostrophe in an EDN stem.
7. Question IDs: SSM uses the printed codice domanda (ssm2025861–ssm20261220); EDN uses constructed IDs (edn2023_P1_KFP1_q2); PES uses <exam_id>_qNNN; MIR and PNA have no stable codes in the PDFs, so question_id is empty there.
8. Coordinator note: the SSM (22) and PNA (3) image files were re-extracted during final assembly after the workers' copies were lost to a VM replacement; filenames match the CSV references exactly and content was spot-verified against the workers' descriptions.

Build artifacts (for re-runs/audits)

• EDN: build_edn_csvs.py, edn_france_build_issues.log (this folder)
• PES: ~/workspace/pes_parse/ (parsers, OCR caches, manual_keys/*.json)
• MIR: /tmp/mirbuild/ (ephemeral — may be gone)