Transmission, decomposed (stage B4, section 3)

Computed at B4a-01, before any ruling touched assign_final.jsonl (out/_analysis/b4_prepare.py, b4_transmission.py).

Over which items. The rows exist only where my own answer is on record: the 40 gold items (gold_items.jsonl) and the 50 items stage B2 piloted (author_items.jsonl). My answer is the latest recheck in rechecks-B2.jsonl, which all 90 carry; my confidence is the file's. The readers are stage B3's two corpus replicates of each item; they agree when their sets are equal. For the gold items, the seeded copies' commonest set and its share are given beside, as a second view; the seeded placements enter no rate. Outside the gold and pilot items my answer does not exist, so the split rate there cannot be decomposed, and nothing here is extrapolated from the gold set, which was drawn to be hard.

The next table counts the items in each row, gold and pilot apart.

My answer was 'clear' on 9 of the 90 items (4 gold, 5 pilot) and 'arguable' on 81, so a split mostly lands in row 3, kept as a finding, and row 2 holds one item.

Row 2: fails to transmit (1 items)

The readers split where I was confident: a wording defect, the sentence stage C1 must repair.

Row 3: the instrument does not decide this item (18 items)

The readers split and I was unsure too: the instrument does not decide these items. Kept as a finding, never revised away.

Row 4: teaches something unintended (20 items)

Both readers agreed with each other and not with me. Each is read and split both ways: I was wrong, or the scheme taught something I did not intend.

Row 1: the scheme decides (51 items)

The scheme decides: both readers placed the set I placed.

By family

The next table gives each family's items in rows 2, 3 and 4 (the family of my primary).

What row 4 says

Which way each went. 11 of the 20 went against me: I was wrong (AyyMD/t08, cloud9/t12, cubers/t27, DevonUK/t04, dofus/t01, floridagators/t04, harrypotter/t08, natsume/t01, overwatch/t23, Slovakia/t07, uiuc/t04). 9 went against the scheme: it taught something I did not intend (alberta/t04, finland/t06, Goodboye/t03, ireland/t12, korea/t15, mit/t01, nVade/t04, parabola/t02, uofmn/t01).

Rule 0's added theme did not transmit. On alberta t04, finland t06, korea t15 and mit t01 the primary's own decide_when_both tells the reader to add the second theme 'under rule 0', and both readers placed one theme. Readers apply rule 0's single placement more strictly than the scheme's pairing sentences ask. This is the same lean stage B2 found, and stage C1 should read these four sentences first.

Rule 2 keeps themes both readers left out. 11 of the 20 row-4 items are class 2, so assign_final.jsonl keeps my B1 grouping on them, the readers' list first, then B1's other themes (J, 12 October). On 8 of those (AyyMD/t08, cubers/t27, DevonUK/t04, floridagators/t04, harrypotter/t08, natsume/t01, overwatch/t23, uiuc/t04) I judge the kept theme wrong on reading. The class-2 lines stand as J ruled; this is reported, not overruled, and no agreed item outside these 90 was re-read in search of more.

What this licenses. The readers share an architecture with me and with each other, so these rows bound whether the instrument decides, not whether it is right. A row 1 item says the written scheme reached two readers the way I meant it; it does not say the placement is true of the room.