# Reporting a reproduction or discrepancy

Publish the complete `validation-run.json`, preregistration, residuals and any
modified inputs at a stable public URL. Preserve failed commands and raw output;
do not submit only a pass/fail summary.

Then notify HBF Network through the public project channel listed at
[github.com/hbfnetwork](https://github.com/hbfnetwork). If the relevant source
repository exposes an issue tracker, open an issue there and link the immutable
evidence. If no tracker is available, publish the evidence first and use the
public HBF Network community link on the website to identify its URL. Do not
send credentials, embargoed datasets, personal data, export-controlled material
or unlicensed publication text.

Include:

1. archive and source-tree SHA-256 values;
2. operating system, architecture, Python/Node/NumPy versions;
3. exact command and unedited output;
4. model ID, version, input domain and observable;
5. expected and observed values, units, residual definition and tolerance;
6. whether the deviation reproduces from a fresh extraction;
7. all local patches as a unified diff;
8. conflicts of interest and reviewer identity, if claiming external review.

Classify the report as one of: packaging failure, non-determinism, numerical
disagreement, domain/units defect, provenance/hash defect, claim-boundary
defect, security issue, or proposed external evidence. Security vulnerabilities
should follow the security policy of the source repository rather than being
published with exploit details.

Submitting a report does not automatically change a model's maturity. The
evidence must be reviewed against
`platform/science/validation-evidence.schema.json`, and failures remain part of
the public record.