# Supplement 1

Bolduc É. Reporting of the registered primary outcome in drug-trial press releases: two preregistered cross-sectional studies. PPN World Research Paper No. 3.

**eAppendix 1.** Registration, pilots, deviations and run notes
**eAppendix 2.** Codebook (version 2), used unchanged by both studies
**eTable 1.** Agreement between coders
**eTable 2.** Robustness analyses
**eTable 3.** Study 2 by filing year (descriptive)
**eTable 4.** Study 2 documents: classification and losses
**eTable 5.** Exploratory: filing time of first announcements (study 2)
**eTable 6.** Cost of the language-model work

---

## eAppendix 1. Registration, Pilots, Deviations and Run Notes

**Study 1** was registered on November 11, 2026 (repository commit 15dc3a17): plan, codebook version 2, analysis code, the pilot's 2 agreement tables, and a manifest of their SHA-256 hashes timestamped with OpenTimestamps. Before registration, a pilot of 823 English-language releases published October 7 to November 8, 2026 (after the frame's window), 199 of which named a ClinicalTrials.gov identifier, was fetched and coded twice; its results were seen and changed the design in 3 ways, all disclosed in the registration: the main analysis was restricted to randomized trials with an efficacy primary outcome, releases not issued by the trial's side were excluded, and the registry record was taken as it stood on the release date. The pilot's first coding round used OpenAI's GPT-5.5 as second coder; the author then chose Claude models only, and the registered check coder is Claude Opus 5.5.

*Deviation 1 (November 11, 2026).* Robustness analysis R6 split releases by publisher type with a label ("distributor") that the source census does not use ("wire"), so its distributor half was empty. The label was corrected; no main estimate or hypothesis reads it. The output of the registered code is deposited beside the corrected one.

**Study 2** was registered on November 12, 2026 (commit 9f70ece2), after study 1's registered and exploratory results were known; its hypotheses restate study 1's exploratory findings, with study 1's figures given in the registration. Before registration, a seeded random pilot of 140 candidate documents (seed 20261112) was fetched, classified and coded, and its results were seen; the pilot's documents and any document with the same text are excluded. Claude Haiku 5.5 was chosen as measurement coder after it recoded study 1's 150 double-coded releases (agreement with Opus, Krippendorff α: verdict stated 1.00, headline states the primary result .92, not met 1.00, met .97, first announcement .93, conference or publication .88).

*Deviation 1 (November 12, 2026).* The download script imported its User-Agent from the search script, and importing a script runs it: started in a new folder without the search log, it began the search again and appended 164 lines to the registered candidate list before it was stopped, about 1 minute later, with no document fetched. The list was cut back to its registered 10,791 lines (SHA-256 equal to the manifest's again) and made read-only, and the run's folder linked the complete search log. No registered file changed.

*Run notes (not deviations).* 9,820 documents were fetched; 19 failed on the first pass (18 HTTP 503 responses, 1 network error) and all were fetched on a second run of the unchanged script. The text-extraction script was rerun unchanged with a larger memory limit. 6 documents could not be classified (3 refusals, 3 reached the output limit) and are missing. 84 headlines were not found character for character in the text (mostly quotation marks and capitalization) and were kept, as registered. 24 releases reached the coder's output limit; following the script's own instruction, they were resubmitted once to the same model with the same settings: 2 were coded, and 22 (long filings of 6,394 to 37,851 words) are missing.

## eAppendix 2. Codebook

### Codebook — trial results in press releases (Paper No. 3)

Version 2 (after pilot coding round 1: α below .667 on the verdict, headline
and spin questions; categories that coders split are merged or made binary).

One coding = one (release, trial) pair. The coder sees the release text and the
trial's registry record as it stood on the release date (registered primary and
secondary outcomes, phase, allocation, sponsor class, status). The coder never
sees the publisher's name or website. Every evidence field quotes the release
verbatim (≤ 40 words); "" when the answer is "no" or "n/a".

**Code each trial on its own.** A release often names several trials. Answers
and evidence for a trial must concern THAT trial — never a sister trial, an
earlier phase, or the drug's results in general.

**The registered primary outcome** is the outcome (or outcomes) listed under
REGISTERED PRIMARY OUTCOMES for this trial. The release may use other words for
it (an abbreviation, a responder version of the same measure, the same measure
at the registered time point); it is the same outcome if the measure and the
time point match. A different measure, a different time point, a subgroup or a
post-hoc version is NOT the registered primary outcome.

#### R1. Registered primary outcome type (from the registry record only)
`efficacy` if at least one registered primary outcome measures benefit
(response, survival, symptoms, a disease or biomarker measure used as the
efficacy measure); `safety/PK/dose` if every registered primary outcome is
adverse events, tolerability, laboratory or vital-sign changes,
pharmacokinetics, immunogenicity, dose-limiting toxicity or dose finding.

#### Q0. Who issued the release? (once per release)
- `trial side` — the lead sponsor, a collaborator or licensing partner, or an
  institution of the trial's investigators, speaking in its own name (first
  person, its own quotes, its own contacts).
- `third party` — anyone else: a news-style rewrite of another
  organization's announcement, a law firm's investor alert, an analyst or
  market-research note, a conference organizer's release about many trials.
If `third party`, stop.

#### Q1. Does the release report results of THIS trial?
- `results` — any efficacy or safety finding from this trial, including a
  regulatory approval release that states the trial's findings, and a
  qualitative statement such as "well tolerated" or "showed activity".
- `no` — the trial is only mentioned (start, enrolment, design, expected
  timing, a pipeline list, an approval or filing that cites it without any
  finding, a publication announced without any finding).
If `no`, stop.

#### Q2. What kind of results announcement?
`topline` (first announcement of the main result) · `interim` · `final/full`
· `conference` · `publication` · `regulatory` (approval or filing) · `other`.

#### Q3. Does the release name the trial's registered primary outcome?
`yes` — it names the measure (for a trial with several registered primary
outcomes: at least one of them). `no` — it does not ("efficacy", "safety",
"the primary endpoint" with no measure named do not count).

#### Q4. What does the release say about the result of the registered primary outcome?
Work down the list; take the first that applies.
1. `not met` — says the outcome was not met, missed, did not reach
   statistical significance, or gives a p-value at or above the stated
   threshold (0.05 if none is stated).
2. `mixed` — several registered primary outcomes, at least one met and at
   least one not met.
3. `met` — says it was met, achieved, reached statistical significance, or
   gives a p-value below the threshold — for the registered primary outcome.
   "The study met its primary endpoint" counts as `met` even if the measure is
   not named, PROVIDED the release does not name a different outcome as that
   primary endpoint (then see Q5 and answer `no verdict`).
4. `number, no verdict` — gives a result for the registered primary outcome
   (an effect, a rate, a difference, a count) without saying whether it was met.
5. `no verdict` — anything else: the outcome is not mentioned, or mentioned
   without a result, or only described with words such as "encouraging",
   "clinically meaningful", "well tolerated", "consistent with".

#### Q5. Does the release call some OTHER outcome this trial's primary?
`yes` — the release names, as this trial's primary (or main, principal)
outcome or endpoint, a measure or time point that is NOT a registered primary
outcome. Give the release's words. `no` — otherwise.

#### Q6. Does the HEADLINE state a result of the registered primary outcome?
`yes` — the headline gives this outcome's verdict ("met its primary
endpoint", "missed primary endpoint") or names the outcome with a result.
`no` — otherwise (a generic "positive results", a secondary outcome, a
subgroup, safety, a milestone, an approval).

#### Q7. Does the HEADLINE present the results as favourable?
`yes` — positive, met, superior, significant, improved, durable, promising,
encouraging, or an approval. `no` — neutral or unfavourable.

#### Q8. Spin (only when R1 = `efficacy` and Q4 is `not met`, `mixed`,
`number, no verdict` or `no verdict`). Mark each that applies:
- `other favourable results` — the release highlights favourable results
  OTHER than the registered primary outcome: secondary outcomes, subgroups,
  post-hoc or responder analyses, change from baseline within one arm, an
  open-label extension;
- `trend as benefit` — a difference that is not statistically significant is
  presented as benefit ("trend", "numerically", "nominal", "clinically
  meaningful" without significance);
- `beneficial conclusion` — the release concludes the treatment works or is
  advancing to the next phase or to regulators, on the strength of the above;
- `none`.

#### Q9. If Q4 = `met`: does the release give at least one number for the
registered primary outcome (effect, rate, difference, p-value or interval)?
`yes` · `no`.

#### Q10. Coder's confidence: `high` · `medium` · `low`.

## eTable 1. Agreement Between Coders

Study 1: coder A (Claude Sonnet 5.5) vs coder B (Claude Opus 5.5) on the 150 releases both coded (59 randomized efficacy results pairs). Study 2: coder H (Claude Haiku 5.5) vs coder B (Claude Opus 5.5) on the 199 releases both coded (82 randomized efficacy results pairs). α is Krippendorff's α for nominal data; an α of 0 with high agreement reflects a rare category. A hypothesis whose measures had α below .667 was to be reported as undecided.

| Question or measure | Study 1 n | agreement | α | Study 2 n | agreement | α |
|---|---|---|---|---|---|---|
| Q0 issued by trial side | 150 | 98.0% | .79 | 199 | 99.5% | .00 |
| R1 primary type | 216 | 98.1% | .95 | 358 | 94.4% | .87 |
| Q1 results | 216 | 95.4% | .90 | 358 | 98.0% | .96 |
| Q2 kind | 84 | 85.7% | .83 | 128 | 82.8% | .79 |
| Q3 registered primary named | 84 | 91.7% | .81 | 128 | 88.3% | .76 |
| Q4 result (5 values) | 84 | 84.5% | .77 | 128 | 90.6% | .86 |
| Q5 other outcome called primary | 84 | 96.4% | .65 | 128 | 98.4% | .85 |
| Q6 headline states primary result | 84 | 96.4% | .78 | 128 | 97.7% | .83 |
| Q7 headline favourable | 84 | 72.6% | .42 | 128 | 92.2% | .85 |
| M1 primary named | 59 | 89.8% | .73 | 82 | 92.7% | .84 |
| M2 verdict stated (met / not met / mixed) | 59 | 98.3% | .97 | 82 | 95.1% | .90 |
| M3 outcome switched | 59 | 98.3% | .00 | 82 | 98.8% | .85 |
| M4 any spin (no-verdict/not-met pairs) | 30 | 80.0% | .29 | 50 | 90.0% | .70 |
| spin: other favourable results | 30 | 80.0% | .45 | 50 | 96.0% | .90 |
| spin: trend as benefit | 30 | 96.7% | .84 | 50 | 92.0% | .67 |
| spin: beneficial conclusion | 30 | 70.0% | .41 | 50 | 80.0% | .60 |
| M5 headline states primary result | 59 | 94.9% | .77 | 82 | 96.3% | .71 |
| M6 number given when met | 27 | 96.3% | .92 | 28 | 96.4% | .93 |
| not met (Q4 = not met) | 59 | 100.0% | 1.00 | 82 | 100.0% | 1.00 |
| met (Q4 = met) | 59 | 96.6% | .93 | 82 | 95.1% | .90 |
| Q2 topline (H1–H3) | — | — | — | 82 | 97.6% | .91 |
| Q2 conference or publication (H2) | — | — | — | 82 | 90.2% | .78 |

## eTable 2. Robustness Analyses

Each hypothesis's difference, percentage points [90% CI]; ✓ = supported by the registered rule (90% CI excludes 0 in the predicted direction and Holm-adjusted P < .05). None of these sets decides a hypothesis.

**Study 1** (H1, headline states the primary result: met minus not met; H2, verdict stated: large minus small sponsors)

| Set | Main-set pairs | H1 | H2 |
|---|---|---|---|
| R1 double-coded releases, coder A (Sonnet) | 51 | −10.0 [−47.8 to 30.0] | 24.6 [−4.3 to 47.6] |
| R1 double-coded releases, coder B (Opus) | 57 | 14.8 [4.0 to 33.3] ✓ | 21.1 [−10.7 to 48.6] |
| R2 registry record unchanged around the release | 253 | 8.4 [−7.1 to 23.4] | 18.4 [7.2 to 30.1] ✓ |
| R4 one pair per trial (earliest release) | 196 | 6.2 [−11.4 to 23.2] | 17.8 [4.6 to 29.7] ✓ |
| R5 phase 3 trials only | 185 | −3.2 [−24.0 to 17.0] | 0.3 [−13.1 to 15.5] |
| R6 distributors' websites | 248 | 10.7 [−5.9 to 27.1] | 19.1 [7.7 to 31.2] ✓ |
| R6 organizations' own websites | 14 | −10.7 [−54.2 to 25.0] | −27.3 [−42.9 to 0.0] |

Main-set pairs whose trial had results posted on ClinicalTrials.gov (R7, count only): 60.

**Study 2** (H1, first announcements, headline states the primary result: met minus not met; H2, verdict stated: first announcements minus conference or publication releases; H3, "not met" among verdicts: first announcements minus other releases)

| Set | Main-set pairs | H1 | H2 | H3 |
|---|---|---|---|---|
| R1 double-coded releases, coder H (Haiku) | 62 | −40.9 [−100.0 to 18.2] | 45.5 [17.3 to 73.3] ✓ | 6.7 [−12.5 to 30.3] |
| R1 double-coded releases, coder B (Opus) | 65 | 11.1 [0.0 to 40.0] | 69.4 [45.5 to 90.0] ✓ | 8.2 [−15.0 to 37.5] |
| R2 registry record unchanged around the release | 2,017 | 14.3 [3.4 to 23.6] ✓ | 43.3 [36.0 to 51.2] ✓ | 18.5 [13.7 to 22.8] ✓ |
| R4 one pair per trial (earliest release) | 801 | 18.5 [4.7 to 29.4] ✓ | 53.0 [45.4 to 60.9] ✓ | 14.2 [7.4 to 20.6] ✓ |
| R5 phase 3 trials only | 1,290 | 10.5 [−5.8 to 23.7] | 33.8 [23.5 to 45.4] ✓ | 13.3 [8.2 to 17.8] ✓ |
| R6 lead sponsors as clusters | 2,141 | 14.0 [3.7 to 23.7] ✓ | 42.9 [35.2 to 50.6] ✓ | 18.4 [13.7 to 22.9] ✓ |
| R7 Form 8-K filers (domestic) | 1,191 | 2.5 [−7.6 to 13.1] | 51.8 [43.8 to 59.3] ✓ | 17.0 [10.3 to 23.2] ✓ |
| R7 Form 6-K filers (foreign) | 950 | 31.2 [9.1 to 44.6] ✓ | 33.9 [21.8 to 46.3] ✓ | 15.5 [9.6 to 19.9] ✓ |

Main-set pairs whose trial had results posted on ClinicalTrials.gov (R8, count only): 214.

## eTable 3. Study 2 by Filing Year (Descriptive)

| Filing year | Main-set pairs | Verdict stated | Headline states the primary result | First announcements | "Not met" among first-announcement verdicts |
|---|---|---|---|---|---|
| 2017 | 62 | 66% | 16% | 21 | 33% |
| 2018 | 88 | 64% | 13% | 24 | 14% |
| 2019 | 101 | 61% | 11% | 19 | 32% |
| 2020 | 145 | 70% | 17% | 44 | 39% |
| 2021 | 249 | 59% | 12% | 75 | 31% |
| 2022 | 281 | 50% | 9% | 50 | 33% |
| 2023 | 280 | 53% | 15% | 70 | 31% |
| 2024 | 314 | 44% | 14% | 82 | 36% |
| 2025 | 395 | 56% | 12% | 114 | 24% |
| 2026 | 226 | 52% | 9% | 54 | 35% |

## eTable 4. Study 2 Documents: Classification and Losses

| Step | Documents |
|---|---|
| EDGAR full-text search results ("NCT*", Forms 8-K, 6-K, 8-K/A, 6-K/A), unique documents | 10,708 |
| EX-99 exhibits and Form 6-K documents filed January 1, 2017, to June 30, 2026, outside the pilot | 9,820 |
| Fetched | 9,820 |
| Naming a ClinicalTrials.gov identifier | 9,174 |
| Classified: press release | 6,881 |
| Classified: presentation | 1,829 |
| Classified: scientific (poster, abstract, manuscript) | 141 |
| Classified: other (letters, transcripts, report pages) | 317 |
| Not classified (refusal or output limit) | 6 |
| Press releases after duplicates and pilot copies removed | 6,856 |
| … filed on Form 6-K (foreign issuers) | 2,265 |
| … filing companies | 700 |
| Release–trial pairs without a registry record (trial not in the snapshot in force) | 450 of 13,906 |

## eTable 5. Exploratory: Filing Time of First Announcements (Study 2)

EDGAR acceptance time (US Eastern) of the filings carrying main-set first announcements (1,703 main-set filings timed). Shares with 95% CIs from the filer bootstrap. Not prespecified as a test (registration §9: exploratory).

| Filed | Met (n = 291) | Not met (n = 140) |
|---|---|---|
| Before the market opened (before 9:30 AM) | 73.9% (215 of 291; 95% CI, 65.6%-81.3%) | 61.4% (86 of 140; 95% CI, 51.4%-71.5%) |
| After the market closed (4 PM or later, any weekday) | 17.5% (51 of 291; 95% CI, 11.4%-25.7%) | 27.1% (38 of 140; 95% CI, 18.3%-37.5%) |
| On a Friday | 13.7% (40 of 291; 95% CI, 8.8%-18.1%) | 9.3% (13 of 140; 95% CI, 4.4%-14.0%) |
| On a Friday, 4 PM or later | 2.1% (6 of 291; 95% CI, 0.4%-4.2%) | 0.0% (0 of 140; 95% CI, 0.0%-0.0%) |

## eTable 6. Cost of the Language-Model Work

Message Batches API prices (half the list price), from each request's recorded token use.

| Step | Model | Cost, US$ |
|---|---|---|
| Study 1, coder A (every release) | Claude Sonnet 5.5 | $6.73 |
| Study 1, coder B (150 releases) | Claude Opus 5.5 | $2.81 |
| Study 2, classifier (every document) | Claude Haiku 5.5 | $1.58 |
| Study 2, coder H (every release) | Claude Haiku 5.5 | $5.78 |
| Study 2, coder B (200 releases) | Claude Opus 5.5 | $4.48 |
