# Verifier notes (reproduction, job:0f8d59e8fc4e28d856fa9b6b21449e33)

## (a) Integrity flag: "95" in paper.md line 35

The flagged "95" is in "Newcombe 95% intervals" and gives the confidence level of the post hoc intervals. It is not a reported result. code/effects.py computes these intervals with z = 1.959964 in `wilson()` (the two-sided 95% normal quantile), combines them with the Newcombe hybrid-score method in `newcombe()`, and uses `stats.t.ppf(0.975, df)` for the Welch anterior-minus-posterior interval. It writes them under the key `ci95_pp`. So the number matches the code. No finding.

I also checked the numbers written in claims.json statements against the re-run outputs, and they all match. Examples: interoception anterior +26.2 pp; time perception CI -11.7 to 15.33; vestibular CI -15.8 to 5.05; interoception A-P -1.18 (-5.98 to 3.61); vestibular A-P -12.83 (-22.78 to -2.88); time perception A-P CI -18.96 to 18.04; mood anterior +2.78; sadness A-P -1.03. The H3 mean also checks out: the symptom z values -1.071, 0.085, 3.504 and 1.711 average to 1.057.

## (b) Hazard screen

I read every file under bundle/: paper.md, claims.json, deviations.json, materials.json, provenance.json, references.json, env/requirements.txt, data/external.json, code/run, code/convergence.py, code/effects.py, plan/analysis-plan.md, plan/convergence.py, and results/R1-R4.json.
- The content is an aggregate literature meta-analysis (term/coordinate co-occurrence). It contains no CSAM.
- It offers no CBRN or cyber uplift.
- It contains no private data. The inputs are the public NeuroQuery v1 corpus (BSD-3-Clause) and the Harvard-Oxford atlas (FSL), both fetched by hash through data/external.json. The bundle holds no personal records and no credentials.
- It copies no third-party material. The code and text are the authors' own, and the external data is referenced, not included.
- The bundle contains no instructions aimed at verifiers. The harness scan also found no hidden content.

Answer: none.

## (c) Run outcome

The harness (sj-harness 0.8.3) ran in rootless podman 5.4.2 with no network. It fetched and hash-checked all 5 external files, and `sh code/run` exited 0 after 9.19 s (about 27 s wall time including the image step). All 24 declared values reproduced exactly, giving C1, C2 and C3 all reproduced. The re-run results/R1-R4.json are identical to the bundle's results files. code/convergence.py is byte-identical to the registered plan/convergence.py. The deviations listed (atlas archive extraction; post hoc effects.py) agree with the code. Sensitivity runs: the conjunction percentile is 0.4534 at K=1 and 0.5593 at K=3.
