$5,000 clean-room reproduction bounty: Bernstein obstacle and curved Hertz results

Goal

Obtain a genuinely independent third-party reproduction of the corrected computational claims associated with the Bernstein obstacle package.

Corrected frozen target

• Branch: replication/bernstein-obstacle-v2-corrected
• Commit: f2bd41f19ff5afbcca8a23f9afdcdf084364dae4
• Correction PR: #106
• Existing in-project reproduction workflow: Bernstein obstacle full reproduction

The branch is immutable for this replication challenge. The original replication/bernstein-obstacle-v1 branch remains available for provenance but is not the target for new reproductions.

Required clean-room protocol

1. Use a machine/account not controlled by the repository owner.
2. Clone the corrected frozen branch without accepting generated result files as evidence.
3. Run the one-command reproduction workflow or independently reconstruct the same tests.
4. Record OS, Python, BLAS/LAPACK, package versions and hardware.
5. Preserve stdout/stderr, generated CSV/JSON files and SHA-256 digests.
6. Report every tolerance pass/failure, including:
   • simplex certificate stress tests;
   • face-orientation/conformity tests;
   • clipping projection/KKT tests;
   • codimension-one strip-scaling fit;
   • cut-element coefficient localization;
   • curved custom Hertz–Signorini solver;
   • independent scikit-fem assembly inside the supplied project;
   • cross-framework contact-width, pressure and reaction comparisons;
   • neighboring-mesh phase-sensitivity sweep.
7. Independently inspect that the selected-mesh 48.9x width comparison is not treated as a uniform accuracy claim.
8. Post a signed public report with verdict REPRODUCED, REPRODUCED WITH DIFFERENCES, or FAILED.

Acceptance criteria

A valid result must be performed outside this project. Internal reruns, repository-owner actions and automated-agent outputs do not count as third-party replication.

A failure or materially different result is fully eligible if accompanied by complete logs and a reproducible diagnosis.

Proposed bounty

$5,000 after the report and evidence pass public review.

Current comparison values

• custom curved minimum gap coefficient: 0;
• custom curved half-load/reaction error: approximately 1.1e-11;
• maximum custom/scikit half-width difference: approximately 5.7e-6;
• maximum custom/scikit reaction difference: approximately 1.2e-11.

The contact-width estimators are mesh-phase sensitive. These values are comparison targets, not values the reproducer should force.

Clone the immutable branch replication/bernstein-obstacle-v2-corrected at commit f2bd41f19ff5afbcca8a23f9afdcdf084364dae4, then run the Bernstein obstacle full reproduction workflow or independently reconstruct the listed tests. Record the environment, preserve logs and generated files with SHA-256 digests, inspect the mesh-phase comparison, and publish a signed report with one of the required verdicts.