# Deviations from the registration — PPN World Research Paper No. 3, Study 2

Each change made to the registered plan, code or run after registration
(2026-11-12, commit 9f70ece2), numbered, dated and explained.

## Deviation 1 — 2026-11-12: the fetch step re-ran the search; the run procedure changed, not the code

`sec/2-fetch.mjs` takes its User-Agent from `sec/1-search.mjs` by importing it,
and importing a script runs it. In the pilot's folder every month was already
in the search log, so the import sent no request. The main run was started in
a new folder without that log (12:11 UTC): the import began the search again
and appended 164 lines (2016-01 → 2016-07) to the registered candidate list
before it was stopped about a minute later. No document had been fetched.

The candidate list was cut back to its first 10,791 lines; its SHA-256 equals
the manifest's again (`e5f0f210…`). The appended copy is kept
(`candidates.corrupted-2026-11-12.jsonl`). The list is now read-only, and the
main run's folder links to the complete search log, so the import finds every
month done and sends no request. No registered file changed; the frame is the
registered one.

## Run notes — 2026-11-12 (not deviations: no plan, code or frame change)

- **Fetch.** 9,820 documents; 19 failed on the first pass (18 × HTTP 503 from
  EDGAR, 1 network error). Their log lines were set aside
  (`fetched.failed-first-pass.jsonl`) and the unchanged script was run again:
  all 19 came back (HTTP 200). The candidate list kept its registered hash.
- **Extract.** The script ran out of Node's default 4 GB memory; it was run
  again unchanged with `--max-old-space-size=16384`. 9,174 documents name a
  ClinicalTrials.gov identifier.
- **Classify.** 9,168 of 9,174 documents classified; 6 not (3 refusals, 3 hit
  the output cap), recorded as missing. 6,881 press releases; 6,856 after
  duplicates and pilot copies are removed. Headlines not found verbatim in the
  text: 84 (kept, as registered).
- **Registry.** 39 quarterly snapshots, 2017-01-05 → 2026-07-01. 14,793
  release–trial pairs; 464 without a registry record.
- **Coding.** Coder H coded 6,832 of 6,856 releases; 24 stopped at the
  16,000-token output cap. The plan names refusals only; the script's own
  instruction ("re-run to submit those again") was followed once, same model
  and settings: 2 were coded, 22 stayed missing (all long filings, 6,394–37,851
  words). Coder B coded all 200; 199 releases were coded by both.
- **Cost.** Classifier $1.58, coder H $5.59 + $0.19 (the resubmission), coder B
  $4.47 — $11.83 (Message Batches prices).
