An independent rerun begins with a fixed experiment boundary and ends with raw evidence that can be compared to the reported result.
Freeze the claimed experiment boundary
Write the exact result to be checked, including metric, population or test set, model or system, comparator, aggregation, and reported value. Name the repository commit, entry command, configuration, and evaluator expected to produce it. A broad request to reproduce the benchmark is not operational until the target table cell or conclusion has a stable identifier.
Inventory every dependency that can change the outcome: runtime, packages, hardware when relevant, hosted APIs, model identifiers, prompts, seeds, dataset revision, preprocessing, cached artifacts, and environment variables. Record secret names and owners without copying values. Mark dependencies that cannot be frozen so the final disposition does not confuse local repeatability with control over external services.
Start from a clean environment
Build the environment from the supplied manifest in an isolated workspace. Capture resolver output, package versions, hardware and driver details when relevant, data digests, and the exact command. Do not repair undocumented steps silently. Record each change required to make the run execute, because those changes are part of the discrepancy between the artifact and its stated instructions.
ACM artifact-evaluation guidance asks whether artifacts are documented, complete, exercisable, and consistent with the paper or reported result. Use those dimensions separately. An experiment can execute without reproducing the result, and a result can match even when the artifact requires undocumented intervention. The audit should preserve both operational and numerical findings.
Compare raw evidence before conclusions
Retain raw outputs and evaluator logs, then recompute the reported aggregate from those artifacts. Check sample counts, exclusions, failed cases, rounding, seeds, and whether the table used a selected run or an average. Report absolute and relative differences with the buyer's tolerance, but do not let a close aggregate conceal a different underlying sample set or evaluation rule.
Run Record performs the independent attempt through Reality Contact, LLC. The buyer supplies lawful access, names the target claim, and decides how the disposition affects product, technology, operations, procurement, or publication. A passing rerun supports only the named environment, data, evaluator, and tolerance; it does not certify the broader scientific or commercial conclusion.
Where the service stops
Reality Contact, LLC audits technical reproducibility and provenance but does not certify scientific truth, research ethics, statistical validity, security, regulatory compliance, publication acceptance, or fitness for every downstream decision; undisclosed data and inaccessible services remain outside the verdict. The accountable buyer authorizes data and compute access, defines the decision and target claims, resolves missing-evidence questions, and chooses to accept the result, require repairs or reruns, narrow its use, or stop relying on it. This technical reproducibility audit does not replace statistical, scientific, legal, ethics, security, privacy, peer, or publication review. The accountable buyer decides whether to accept, repair, rerun, narrow, or stop relying on each result.
Sources: ACM artifact evaluation guidance; Nature Communications reporting standards.