ilisium
WorkspaceBring your agent
All project proposals

Research

Reproduction Lab

Take one published result apart, reproduce its key experiment, and leave a record another team can inspect.

What the team would deliver.

A runnable reproduction package and an independently reviewed comparison with the original result.

Determine whether one precisely stated published result can be reproduced under documented conditions, including an honest account of discrepancies.

A bounded path to that result.

Suggested work packets for an initial team to scope and review together.

  1. Pin the claim

    Choose one bounded result and agree on the conditions and tolerances that would count as reproduction.

    Output A cited reproduction brief with a fixed evaluation plan.

  2. Reconstruct the experiment

    Resolve setup details, inspect data permissions, and make the smallest runnable implementation.

    Output A pinned environment and executable experiment.

  3. Run and explain

    Record outputs and differences without concealing unsuccessful runs or changing the target after seeing results.

    Output A comparison report linked to raw evidence.

  4. Repeat independently

    Give the package to a reviewer who did not author the implementation and address their findings.

    Output An independent reproduction record and a qualified final conclusion.

Different strengths. Shared work.

People

Subject reviewer
Select a tractable claim and check that the proposed comparison tests what the source actually says.
Experiment lead
Approve the method, data permissions, resource budget, and interpretation of results.
Independent reproducer
Repeat the experiment from the package and challenge unsupported conclusions.

Their AIs

Source analyst
Extract explicit assumptions, method details, and unresolved ambiguities with citations.
Experiment assistant
Prepare scripts, environment manifests, and checks for the agreed experiment.
Comparison assistant
Compare outputs, surface discrepancies, and propose checks without deciding that a claim is proved.

These are proposed responsibilities. People choose their tools, review contributions, and authorize changes; this page does not connect or run an AI.

What would count as done.

  • The team pins one source claim, the relevant method, data, and evaluation target before running the reproduction.
  • A separate reviewer can repeat the documented experiment from the pinned materials.
  • The report distinguishes a reproduced result, a qualified discrepancy, and an inconclusive attempt using the agreed criteria.
  • Negative findings and failed runs remain in the record; agreement with the source is not required for a useful result.

What the record would contain.

  • Source citation, exact claim, and a method comparison approved before the run.
  • Data provenance, environment manifest, parameters, seeds where applicable, and source revision.
  • Raw outputs and logs for successful and failed attempts.
  • Independent rerun report and a human decision stating the limits of the finding.

Help shape the brief.

Explain what you would contribute, what needs to change, or what would make this outcome useful. Scope and participation would be agreed before work begins.

Discuss Reproduction Lab

Opens your mail app. Sending an email does not enroll you or reserve an assignment.