NCT04730349 — SoA extraction uncertainty report

Prompt: PDF_TO_JSON_PROMPT v3.8.1, run in one pass without stopping. Source: NCT04730349_soa.pdf (16 PDF pages = document pages 27–42, per PAGEMAP.md). No protocol markdown was provided, so all text comes from the PDF text layer and was checked against rendered pages.

Decisions needed (3)

#	where	call made	alternative	detail
D1	p.31, Table 2-2 row 4 (Targeted Physical Examination…), Day 8 cell (col 5), marker `*`	The Notes-cell line "At Day 8 visits collect vital signs only." is its own annotation ``, bound to the Day 8 cell that shows "X*". The rest of that Notes cell is annotation n1, bound to the activity.	Keep the whole Notes cell, including the `` line, as one activity-level note (n1). The `` would then stay only as a mark on the cell.	3.3
D2	p.37, Table 2-2 row 29 (Oral Hydration Follow-up), col 4	"X (Day 3-5)" is recorded only in the Cycle 1 Day 5 column (col 4), where it is printed. The rule line between Day 3 and Day 5 is present in that row, so this is not a merged cell. The qualifier stays in the cell text.	Treat "(Day 3-5)" as naming the Cycle 1 Day 3 and Day 5 columns, and spread the X over cols 3:4.	3.2
D3	p.31, Table 2-2 header rows 1–2	The cycle-length lines ("Cycle = 3 wks", "Each cycle = 3 wks") and the visit windows ("± 1 day", "- 1 day") stay inside the cycle and day header cells. This gives two header properties (cycle, study_day).	Split them out into separate header properties (a cycle-length row and a `window` row).	3.1

Recorded, not open (7)

All three tables typed main_soa. Rule: taxonomy "multiple independent schedules (e.g. Screening table and Treatment table with different column structures) → each main_soa". Section 2 (p.27) presents Tables 2-1, 2-2 and 2-3 as three schedules (Screening, On Treatment, Long-term Follow-up). Each has its own column set, and all three cover the same participants.
Each table has one label column (Procedure), so L = 1 and the first data column is position 2. The right-hand Notes column is not a schedule column (§5, §6).
Each non-empty Notes cell is one annotation, bounded by rule lines, with a synthesised marker (n1, n2, …) bound by method: synthesized to the activity row beside it (§6).
Merged text cells are distributed over their span with source_range (§5). This covers "Continuously", imaging instruction text, "Performed … as clinically indicated", "See Section 9.5.1 …", "See Notes" and "As clinically indicated".
Qualified marks are kept as literal cell values because the qualifier is a condition, not a column span (§5): "X (Cycle 5 only)" (T2 row 5, col 7), "If toxicities are present." / "If toxicities are present" (T3 row 10, cols 3–4).
Footnote c of Table 2-3 is printed on the Notes column header. It is treated as table-scope: one schedule_property location with method: synthesized, and no marker placed on any element (§6).
Abbreviation blocks (pp.30, 37, 42) produced no annotations. None of their terms is printed as a marker (§6).

1. Per-table summary

Table	Source	type	pages	data cols	activities (incl. headers)	activity rows per page
1	Table 2-1 Screening Procedural Outline	main_soa	28–30	1 (pos 2)	21	28: 12 · 29: 6 · 30: 3
2	Table 2-2 On-treatment Procedural Outline	main_soa	31–38	6 (pos 2–7)	27	31: 4 · 32: 4 · 33: 2 · 34: 4 · 35: 5 · 36: 6 · 37: 2 · 38: 0
3	Table 2-3 Long-term Follow-up Period	main_soa	39–42	4 (pos 2–5)	19	39: 7 · 40: 3 · 41: 3 · 42: 6

Page 27 holds only the Section 2 introduction, so no table starts there. Table 1 starts on page 28.
Page 38 contributes no activity rows. It holds only footnotes d and e of Table 2-2, and that is expected.
Section-header rows (bold, no marks) are at indentation 0 and procedures at 1 (indentation_method: font_signal). Table 1 headers: Eligibility Assessments, Safety Assessments, Laboratory Tests, Tumor Assessment. Table 2 headers: Safety, Laboratory Tests, Efficacy, PK/Immunogenicity (marker e), Health Outcomes, Study Drug. Table 3 headers: Safety, Laboratory Tests, Efficacy, PK/Immunogenicity.

3.1 Header structure (Table 2)

Row 1 is cycle: "Cycle 1 Only" is merged over 2:5 with markers a,b, and "Cycle 2 and Beyond" is merged over 6:7 with markers a,b,c.
Row 2 is study_day: Day 1, Day 3 (± 1 day) with marker d, Day 5 (± 1 day), Day 8 (- 1 day), Day 1, Day 3-5.
Both property names are synthesised ("Cycle", "Day") because the label cell is the merged "Procedure" heading.
Tables 1 and 3 each have one visit header row with the name "Visit" synthesised.
The Survival Follow-up window "(± 14 Days)" is kept in its cell value.
Header footnote markers are on the specific schedule_grid cells. Their locations are schedule_property entries with a column_position.

3.2 Merged-mark decisions

The spans were confirmed from 200-dpi raster rule lines: the internal vertical rules are missing across each span. - Table 2, span 3:7 (the Cycle 1 Day 1 column stays empty, and its rule line is present): row 6 AE Assessment ("Continuously"), row 7 Concomitant Medication Use ("Continuously"), row 12 Body Imaging (bullet text), row 13 Brain Imaging (bullet text), row 14 CSF - Solid tumors, row 15 Bone Marrow- Solid tumors, row 16 CSF - Leukemia, row 17 Bone Marrow - Leukemia, rows 19–21 PK/Immunogenicity samples ("See Section 9.5.1 for further details."), row 23 PRO ("See Notes"). - Table 3, span 2:5: row 13 Body Imaging, row 14 Brain Imaging, row 15 CSF ("As clinically indicated."), row 16 Bone marrow aspirate ("As clinically indicated"), rows 19–20 PK and Immunogenicity samples. - Oral Hydration Follow-up (T2 row 29): every rule line is present, so the mark is a single cell in col 4 (see D2).

3.3 Synthesised items

Property names "Visit" (T1, T3), "Cycle" and "Day" (T2).
Notes-column markers: T1 n1–n17, T2 n1–n17, T3 n1–n12.
For D1, the * note text was taken out of the Notes cell of T2 row 4. The * marker itself is printed in the grid ("X*"), so its location is not synthesised.

4. Mechanical mark-check

Every SoA page has a text layer and vector rule lines. The text layer is not glyph-spread, so no word reconstruction was needed.
Marks: pdftotext -bbox tokens matching ^[Xx][*a-zA-Z0-9]?$ were binned to the nearest header column centre on every page. The resulting matrix matched the visual read cell for cell, including "X*" (T2 row 4, Day 8) and the "X" parts of "X (Cycle 5 only)" and "X (Day 3-5)". There were no disagreements.
Row and cell bands were recovered from raster rule lines, as in §1d, to confirm spans and Notes-cell extents.

5. Annotation text integrity

Containment pairs were checked against the page, and both are source-faithful. They are separate Notes cells in the source, not one cell split across rows:
T2 n4 "Record at each visit." (Concomitant Medication Use, p.32) is contained in T2 n3 (AE Assessment, p.31).
T3 n8 "See Section 9.1.1 for further details." (Body Imaging) is contained in T3 n9 (Brain Imaging, p.41).
Redacted content was transcribed up to the black box, with "[remainder redacted in source]" appended:
Table 1 notes n14–n17.
Table 2 notes n9–n12.
T2 footnote e has an inline redaction, written as "[redacted in source]".
The superscript in "[18F]fluorodeoxyglucose" was flattened to "[18F]". Bullets are rendered as "•" on separate lines. "±" was restored from the rendered page because the text layer maps it to a private-use glyph.
The label of T2 row 23 is printed in grey font ("Patient-reported Outcomes (PRO) Version of the CTCAE"). It is kept as a normal activity, and the grey is recorded in font_info.

6. Source defects / orphan risk

Redaction blocks that may conceal activity rows:
T1: a full-width block below the last row on p.30.
T2: a block covering the table bottom on p.35 (roughly 5 rows' height), and a full-width band just under the header on p.36.
T3: a full-width block below Pregnancy Test on p.40.
No rows were emitted for any of these regions. The reviewer should confirm against an unredacted source if one is available.
Every annotation has at least one marker location, and no marker is used without a definition. T3 marker c is intentionally not placed on any element (table-scope).

7. Method provenance

activity_name_source.indentation_method: font_signal on all activities (bold headers = level 0).
marker_locations[].method: synthesized on all Notes-column annotations (n*) and on T3 footnote c.
No unresolved locations. No proximity_bounded, raster_band_cells or visual_transcription text.