Statistics review of the E3 tables and text

Reviewer: stats-reviewer subagent, read-only, 16 Oct 2026, on reports/tables/e3_*.csv, reports/figures/e3_selection.png, CLAIMS E-40 to E-47, chapter 7 section 7.5, the README and chapter 8, against docs/prereg/addendum-v2.md and docs/PREREG.md. Every number in section 7.5 and the README matched the CSVs; the findings concern wording, framing and traceability.

#	Sev.	Finding	Disposition
1	major	The LAMDA five-seed rows are the Main-cell contrasts of the addendum (seeds 5 to 9 are conditional), but the table code labels them seed_robustness and lists 10-seed rows as not_computed; with five paired seeds the smallest Wilcoxon p is 0.0625, so no LAMDA row can pass Holm and the only Holm-significant rows are the BODMAS ones, which do not test shrinkage.	Text corrected (7.5, README, CLAIMS E-41, E-42): says the rows are the Main-cell contrasts, that the verdicts on LAMDA rest on the interval rule alone, and that no RQ3 result on LAMDA is supported by an adjusted test. The role label in e3_tables.py is left as is so the committed tables keep their hashes; it is named in the text and in E-42.
2	major	“Positive in all eight rows” read as eight replications; the rows share the random arm, windows and seeds, and LAMDA has about four independent blocks.	Text corrected: not independent, intervals likely too narrow, lowest ends (0.0008 to 0.0024) not to be read as excluding zero with confidence; “in this sample”.
3	major	“Equivalent to zero” misreads the verdict, whose margin (0.01) is as large as most delay-0 advantages.	Text corrected: defined as “interval includes zero and lies within plus or minus 0.01”; LR credibility at 2% added; the weakness of these verdicts stated.
4	major	“Worth up to hundredths of A-AUT on BODMAS” overstated and causal; one lower end is borderline.	Text corrected to “higher by 0.003 to 0.008, interval upper ends up to 0.012”; borderline row and bootstrap-mean note added.
5	major	Score comparison used verdict words without intervals.	Text and heading say point estimates only; “indistinguishable” removed.
6	major	The trap arm’s coverage and credibility are confounded by not retraining, too.	Text (7.5), CLAIMS E-45 and the discussion corrected: nothing is read from the arm as evidence about contamination.
7	major	Class-prior confound understated; the centroid arm (class mix near random) is not above random; no claim that credibility adds over uncertainty.	Added to 7.5 and README, with “no interval computed”.
8	minor	Ablation wording causal and trend-based; “slightly below nominal in most rows” was wrong (all twelve are below 0.90).	Corrected.
9	minor	LAMDA label counts and plain-unit effects missing.	Label counts added (43.6 and 87.3 per slot). The per-slot-of-delay unit for the interaction is not computed; listed as not done.
10	minor	CLAIMS E-44 and E-47 notes (zero by construction; duplicate BODMAS seed sets); E-41 Holm p tests delay-0 advantage.	Fixed in the rows.
11	minor	Figure: connecting line implies a decay shape; footer.	Line removed, footer says “two delays only”; figure regenerated (tables unchanged). The five-seed set is what the figure plots, as its footer says.
12	minor	Chapter 6 section 6.11 was stale.	Updated.

Verdict (reviewer): the numbers and decision words reproduce from the stored tables; the main exposure was interpretive (the LAMDA rows labelled as robustness, verdict words used against a margin as large as the effects). Items 1 to 7 were to be fixed before the text is cited; they are. No third pass has been run on the revised text.

Update, 20 Oct 2026: LAMDA seeds 5 to 9 were run on the researcher’s go, so finding 1 no longer applies to the Main cells; the ten-seed LAMDA rows are the registered primary rows (CLAIMS E-42, section 7.5 rewritten). The text of that paragraph was revised after this review and has not had a further independent pass.