Reproduce

Reproduce

A reproduction is a measurement of a claim. Its value comes from deciding, before anything runs, exactly what number would count as success, and from writing down exactly what was run so the next person can measure it again. A reproduction that adjusts its target after seeing the result is not one.

Non-negotiables

1. Freeze the target first. The claim, the metric and its direction, the expected value with its uncertainty, the tolerance, and the dataset, split and conditions it applies to, written down before execution. If the source gives no uncertainty, choose a tolerance and say why.
2. Canonical path before substitutes. Run the authors' code, config, checkpoint and data as released. Only when that path is unavailable or broken, build a replacement, and mark the result as a reimplementation.
3. Capture everything that determines the number. Code revision (and whether the tree was dirty), data identity and version, environment and package versions, hardware, seeds, the exact command, elapsed time, and every warning.
4. Isolate and preserve. New outputs go to a fresh directory; existing outputs are never overwritten. Raw logs stay separate from derived summaries.
5. Do not fix the result. A failing reproduction is a finding. Debugging is allowed until the canonical path runs as intended; changing hyperparameters, filters or metrics to get closer to the paper is a new experiment, reported as such.
6. One verdict, from the fixed set.

Workflow

• Resolve the exact target and write the admission record.
• Recover inputs: code, data, environment, configuration, checkpoint.
• Run the canonical path (compute_job for anything longer than a few minutes).
• Capture outputs; compute the metric the frozen way; compare with tolerance.
• Verdict and record.

Admission record (write it to reproduction/<name>/target.md):

Claim: <sentence from the source, quoted>            Source: <doi/arXiv/section/table>
Metric: <name>, <direction>   Expected: <value ± uncertainty or interval>   Tolerance: <value, reason>
Data: <name, version, split, filters, license/access>   Conditions: <seeds, budget, hardware>
Canonical path: <repo, commit, command>               Missing: <what could not be recovered>

Inputs. Pin the repository commit named in the paper or its release tag; download data by the authors' script; build the environment from their lock file where one exists and record what changed when it does not resolve. Missing inputs are listed in the record, not silently substituted.

Execution. Runs longer than a few minutes go through compute_job with a name that carries the target, so they are tracked and can be resumed after an interruption. Log the exact command and its exit code. For stochastic results run the seed count the paper used (or at least three) and report the spread. Track metrics with openscience_track when the script permits, so the curves survive.

Comparison. Compute the metric with the frozen definition, from the raw outputs, in a fresh process. Report the reproduced value with its spread, the expected value, the absolute and relative difference, and whether it lies within tolerance. Check the checks that apply: leakage-safe split, checkpoint actually loaded, units, the same evaluation subset.

Verdict

Write reproduction/<name>/report.md: the admission record, the environment capture, the command log, the results table, the verdict, and what a reader should try next. A reimplementation says so in its first line.

Before you hand it over

• The target was written before the run and was not edited after.
• Every number in the report traces to a raw output file in the reproduction directory.
• The verdict is one of the five, with the difference and tolerance shown.
• Untested claims and unrecoverable inputs are listed, not implied fine.