# Reproduction access limitation and independent table checks

The default reference-harness reproduction could not reach the computation stage because the exact pinned flare catalogue, data/meta/gunther2020_table1.dat.gz, was inaccessible. Every formal claim remains **could_not_run**, not mismatch or reproduced. No submitted scientific code ran on the host.

I read the paper, all seven claims, plan, active code and its byte-identical registered copies, the exploratory addition, environment, provenance/materials/references, declared JSON results, and the structures and values of the supplied sample/event/injection tables. The bundle contains astronomical observations/synthetic injections and original analysis code. I found no hazardous capability, private human data or substantial unlicensed third-party copy requiring a finding. The pointed-at external catalogue is not copied into the submitted bundle. Hazard screen: none.

## Source access work

The first target-list request failed TLS chain validation. The public MIT host supplied an old intermediate beside a newer leaf. I obtained the missing public issuer chain from its HTTPS AIA repository, verified it against an already trusted certificate-authority root, retained hostname verification, and fetched the target list with that complete verified chain only for this connection. Its exact declared length and SHA-256 matched. No system trust settings were changed and no certificate checks were disabled.

The MAST script and all 340 NASA light curves were fetched and verified by byte count and SHA-256. The attached file-check record therefore verifies 342 of 343 external files, totaling 685,257,358 bytes. Only the 450,571-byte compressed flare catalogue remains missing.

The original CDS catalogue archive timed out. Its reachable VizieR front end returned a service-unavailable page. The [official CDS status page](https://cds.unistra.fr/) states that a cybersecurity incident has caused some Strasbourg systems to be taken offline and lists alternate mirrors. I tried the official Harvard and IDIA file/download routes; no byte-identical catalogue was obtained (404s or timeouts). The Japanese endpoint refused a connection and the Chinese endpoint was unreachable. These are source-access observations, not evidence against the scientific claims. I did not substitute transformed catalogue values or relax the declared hash.

## Separate arithmetic check

The attached original standard-library script imports no submitted modules and reads only the provided output CSVs. It independently obtains:

- 150 held-out stars and 7,500 held-out injections, within 17,000 total supplied injections.
- At the stated matched false-alarm rate, 4,230 baseline recoveries and 4,180 controlled recoveries, difference -50/7,500 = -0.0066667; 716 baseline-only and 666 control-only recoveries. Thresholds match the declared values to their displayed precision.
- At the reported tuning thresholds, 1,067 and 2,446 held-out recoveries, with zero observed null events in both arms.
- The identified tuning star contributes 19 of 22 S/N-at-least-ten null events, median spacing 1.389593 days, all with 24-hour best durations.
- All six reported point differences across false-alarm rates: 0.0788, 0.0660, -0.0066667, -0.0182667, -0.0701333 and -0.0974667.
- For every supplied injection, vetoes never increase the best candidate S/N before thresholding.

Those table-level agreements do not regenerate raw-flux processing, event features, the sample-selection catalogue check, tuned controls or bootstrap intervals. They do not qualify as a complete reproduction from the declared external inputs. The reference-harness proposals are therefore retained as could_not_run for C1-C7. Cached hash-verified files remain available for a future run when the exact catalogue is accessible.
