Use this before submission and again before camera-ready. At FAccT, transparency is not only the subject of the field — it is a norm the community holds its own papers to. But FAccT reproducibility is broader than "does the code run": it spans releasing and documenting the data and models behind an audit, making a qualitative study auditable without exposing participants, and being honest where confidentiality or proprietary access genuinely bars release. The goal is that a competent reader could trace how you got from evidence to conclusion — and judge whether the harm you claim is real.
| Claim in the paper | Weak availability answer | FAccT-ready answer |
|---|---|---|
| "We audit N deployed systems" | "Data available on request" | Released dataset (or documented access) + analysis code + subgroup definitions |
| "Our dataset is representative" | Raw files with no context | A datasheet: how collected, who is in it, gaps, intended and off-label uses |
| "Our model behaves fairly" | Weights only | A model card: evaluation disaggregated by group, intended use, known failure groups |
| "We interviewed P affected people" | Nothing (privacy cited vaguely) | Protocol + codebook + aggregate results + a clear, specific confidentiality boundary |
| "The LLM produced these outputs" | "We used a chatbot" | Model IDs and dates, prompts, cached raw outputs, sampling settings |
"Available on request" reads as not available; convert every such line into a concrete release, proper documentation, or an explicit, justified exception.
[Scraped/mined data] record source, extraction date, and terms; archive the extracted dataset,
not just the scraper; document deduplication and filtering
[Protected attributes] document how group labels were obtained/inferred and their error
[Models] record exact model identifiers + access dates; cache raw prompts and outputs;
report sampling settings; a live-API-only study re-samples, it does not reproduce
[Qualitative] version the codebook; log coding decisions; keep an audit trail a second
reader could follow
[Consent] keep the consent/ethics record aligned with what you release
For FAccT, aim turnkey for anything a reviewer could rerun quickly (a fairness-metric recomputation, a plot from released results); confidential interview data or proprietary system access stays documented with the boundary stated. Stating the achieved level honestly beats promising turnkey behavior that fails.
Consider a study combining a quantitative audit of a benefits system with interviews of claimants. Its transparency spine: the audit code with pinned data versions and subgroup definitions; the released (or access-documented) audit dataset with a datasheet; the interview protocol, codebook, and aggregate themes; the consent and ethics record; and one honest paragraph on what cannot be shared (claimant identities, the agency's internal data) and why — so the audit is falsifiable and the qualitative reasoning is auditable, without re-harming participants.
[Finding inventory] <finding -> evidence location>
[Availability] concrete release / documented access / vague / missing
[Documentation] <datasheet / model card / data statement present where relevant? yes/no>
[Provenance gaps] <scrape terms / proxy labels / model caching / codebook>
[Reproducibility level] turnkey / scripted / documented, stated honestly
[Paper fixes] <must appear in the PDF>
[Artifact fixes] <additions before upload, kept anonymous>