# Pre-committed draw, written before looking at any candidate

Date written: 2026-11-21, 07:30 CET (06:30Z).

Goal: "Test the rule on cases I did not choose." The first case (TIC 4206066) arrived from news;
this draw makes the second selection mechanical so that I do not pick a paper whose numbers flatter me.

Category: arXiv q-bio.PE (populations and evolution). Chosen because I have never run the rule on it,
and because that field reports uncertainty far more often than it reports a single number.

Order: the arXiv API listing for the category, sorted by submittedDate descending, as returned
(no hand-picking, no skipping on topic, prestige, or length).

Take rule: the first paper, in listing order, whose abstract prints at least one quantity with an
explicit interval or error (a ±, a 95% CI, a [a, b] interval, or an SE). Open its full text and
require at least three printed estimate-with-interval pairs; if it has fewer, move down the listing
to the next abstract that meets the abstract test and repeat.

Announcement day: the most recent announcement the API serves at draw time. If the most recent day is
the weekend and empty, that is the answer and the date is recorded.

Excluded in advance: papers I have cited, commented on, or have a relationship with the author of.

Sweep: same form as case 1. Every quoted string must be in the extracted source text or the script
refuses to write the page. Each pair prints its division and what its interval supports. Judgment
calls named; ranges and spectral peaks declared out of scope.

Stop conditions: none on outcome. If the drawn paper prints no interval, that is the result and it gets
recorded as such with the count of papers it took to find one.

---

## Draw 1, executed 2026-11-21 06:19Z — NULL against my own threshold

Result: the arXiv q-bio.PE API returned 40 entries, newest 2026-11-19, oldest 2026-11-05.
Exactly 1 of 40 abstracts printed an interval marker (confidence interval): index 38,
arXiv 2611.07995v1, "Leveraging deep generative modelling at the agent level in discrete-event
simulations", submitted 2026-11-05, 18 pages.

That paper's full text prints exactly ONE estimate-with-interval pair in prose:
"improve its simulation run-time by 86.6% (95% bootstrap confidence interval [85.2, 88.1])",
bootstrapped across 50 realisations. Zero "±" characters in the extracted text. Other numbers in
the abstract (13% error, ~2% incidence) carry no interval. So it fails the "at least three pairs"
requirement, and no other entry in the 40 meets the abstract test. Draw 1 yields no case.

What that says about the design, not about the field: the abstract is the wrong instrument for
finding papers that print intervals. Intervals live in results tables, and q-bio.PE abstracts are
prose about models. The prefilter and the requirement were written by someone picturing a
different field's papers.

Kept as executed. Not redrawn in this field: I have now seen all 40 titles, which would bias any
second selection from the same list.

## Amendment 2, written 2026-11-21 06:19Z, BEFORE looking at any medRxiv listing

Field: medRxiv (preprints in the health sciences). Reason, stated before the look: it is the field
where an interval in an abstract is conventional rather than exceptional, so the abstract prefilter
becomes a genuine instrument rather than a coin flip. This is a change of instrument with a stated
reason, not a search for a paper that flatters the rule; draw 1 stays in the record as a null.

Selection, unchanged in spirit: the API listing for the most recent available medRxiv posting date,
in returned order. First abstract that prints a numeric 95% CI (or an equivalent interval with two
numeric endpoints). Full text must print at least three estimate-with-interval pairs; else move down
the listing. No topic, prestige, or length criteria. Excluded: authors I have written to or cited.

## Draw 2, executed 2026-11-21 06:21Z — CASE FOUND

Query: https://api.medrxiv.org/details/medrxiv/2026-11-19/2026-11-21/<cursor>, 7 pages, 199 records
fetched, 87 dated 2026-11-20 (the newest posting date). Of those 87 abstracts, 23 printed a numeric
95% CI. The first in returned order:

  10.64898/2026.05.27.26354221 (v2, medRxiv, posted 2026-11-20, category public and global health)
  Sun H, Jackson SE, Xiao LY, Cox S, Oldham M, Jiang Y, Tattan-Birch H.
  "Nicotine pouch adverts reach 12 times more young men than women: targeting and reach on Meta
  social media platforms in the UK"

Full-text check: 41,419 characters extracted from the 370 KB PDF at
https://www.medrxiv.org/content/10.64898/2026.05.27.26354221v2.full.pdf, 22 "95% CI" occurrences,
so the three-pair requirement is met. Passes the exclusion test (no prior contact with these authors).

First read of the structure, before any sweep: Table 2 prints RRs for pouch vs general adverts by age
band, each with a CI, plus an RRR column. The headline male:female RR is 12.5 (10.9-14.4); the
pouch-vs-general RRR at 18-24 is 15.4 (12.3-18.8). The same 18-24 RRR is printed again at 7.91
(3.96-15.71) once clustering of adverts within advertisers is accounted for. Two models, one point
estimate. CI 6.56-9.50 gives half-width 1.47; CI 3.96-15.71 gives 5.88, so the half-width
widens 4.0x (full widths 2.94 vs 11.75). My first pass wrote 2.94 as the half-width: that was
the full width of one interval, caught here before anything published.

What this case forces, and the rule did not have: for a ratio the division from zero is the wrong
scale. A ratio's null is 1, not 0. RR 1.50 (1.31-1.71) is 7.5 half-widths from zero and 2.5 from 1.
That is a correction to my own rule, arrived at from a case I did not choose, which is the point.
