# Verifier notes

Conflict of interest: the job's COI screen flagged CONFLICT for this operator. The operator worked it under a standing approval to take conflicted jobs. Readers should give this reproduction the weight a conflicted check deserves.

## Run
- sj-harness 0.8.2, rootless podman 5.4.2, container with no network. Image built from env/Dockerfile (python:3.12-slim pinned by digest; requirements pinned). `sh code/run` exited 0 after 13.6 s (whole harness run 88 s including fetches and image build).
- All 197 external pointers (195 Sector 1 SPOC FITS light curves, the MAST curl manifest and the Flatiron mirror of the MIT target list) were fetched by the harness outside the container. Each matched its declared size and SHA-256.
- verify_sample.py rebuilt the selection from the source files: 195 screened, 100 selected, and the sample and the 40/60 split match data/sample.json. sanity.py passed every assertion.
- All 14 declared results for C1 agree within tolerance. Most agree exactly, including every integer count, the difference of 5.5556 pp and both bootstrap CI ends (-10.5556 and 26.6667).
- Regenerated files compared with the bundle's copies: R2.json, sanity.json and bootstrap_difference_pp.json are byte-identical. In R1.json, target_scores.json and injections.json every integer, label and selection is identical, and floats differ only at the floating-point noise level (max relative difference 8e-8 over 6,923 injection fields; 2e-11 in R1). recovery.png differs in bytes only.

## Check on the rounded white-dwarf constants (report only)
evaluate.py uses R_WD = 0.0112 R_sun [(1.44/m)^(2/3) - (m/1.44)^(2/3)]^(1/2). Nauenberg (1972) Eq. 27 gives 0.01125 R_sun and M_ch = 1.454 (mu_e = 2). I re-ran evaluate.py and diagnostics.py with only those two constants changed (evidence/wdconst_evaluate_variant.py), in the same image under rootless podman with no network.
- Peak pulse amplitudes fall by at most 3.9% (the 0.3 M_sun, 100 d, smallest hosts case). Most changes are under 1%.
- No C1 result changes: baseline 239/540, control_1.0 269/540, background counts 6/6, selected penalty 1.0, difference 5.5556 pp, CI high 26.6667. The CI low moves from -10.5556 to -10.5602 pp, which still displays as -10.6.
- Elsewhere a few counts move by 1–2: development baseline 186→185 and control_0.5 183→182; evaluation control_0.25 274 vs 276 (matched) and 196 vs 197 (fixed); baseline at the 20% operating point 344→343. In the evaluation split, 1 baseline injection and 0 selected-control injections flip.
- Conclusion: the rounding does not affect the claim's conclusions.

## Observations (not mismatches)
- The paper's deviations already disclose post-result diagnostic bins, the smoke-test bound change (2500→2700 ppm) and the target-list mirror.
- Thresholds are matched on the evaluation backgrounds. As the paper states, this is a same-rate descriptive comparison and not a deployment calibration.

## Hazard screen
none. The bundle holds public TESS mission data pointers, catalog stellar parameters for public targets, and numerical Python code. It has no private data, no credentials and no copied third-party content (the FITS files are external pointers). No hidden instructions to verifiers were found.