A run manifest makes the exact experiment invocable while its receipts show what environment and inputs were actually used.
Identify source and executable entry points
Record the repository, immutable commit, subdirectory, entry command, configuration file, and any patch applied after the commit. Include the build or container definition and lockfile. If generated code or notebooks participate, name the execution order and exported script or notebook version. A README description is useful, but the manifest should point directly to machine-checkable identifiers.
Nature's code-publication guidance asks authors to document code versions, dependencies, test data, parameters, and instructions needed to run the code. For a private internal audit, the same fields can remain in controlled storage. Reproducibility does not require making confidential assets public; it requires that authorized reviewers can identify and execute the governed artifact.
Give data and model inputs receipts
Name the dataset, revision, acquisition date, source, license or access rule, preprocessing code, filtering, split definition, and a digest for the exact files or manifest. For hosted or too-large data, preserve a query, snapshot identifier, or provider receipt. Record model names and revisions separately from local checkpoint digests, especially when a hosted alias can change without a repository commit.
Prompts, templates, few-shot examples, tool definitions, and evaluator instructions are experiment inputs. Store them under versions or digests and identify any runtime assembly logic. Secret values remain outside the manifest, but their presence and role should be declared. An environment that runs with a different private endpoint, region, or feature flag is a different experimental condition.
Capture outputs and external uncertainty
Define raw-output, log, intermediate-artifact, and summary-table paths before the run. Record exit status, timestamps, runtime, failed cases, output digests, and the script that aggregates results. Keep enough raw evidence to audit the table without rerunning an expensive or unavailable service, subject to the buyer's data and retention policies.
Run Record builds the manifest through Reality Contact, LLC and marks each field supplied, derived, missing, mutable, or not applicable. The buyer approves retention, access, and the permitted experiment boundary. A complete manifest improves reconstruction but cannot freeze provider-managed models, undisclosed platform changes, or external data sources that offer no stable revision.
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: Nature guidelines for code publication; Weights & Biases experiment tracking product information.