AI-enabled work in research organizations: preserve the path from source to finding

Start with a lineage record, not a prose summary

Create one record for every finding that might influence a conclusion. Give it a stable finding ID and capture the source dataset or document, version, access date, permissions, extraction method, transformations, analysis code or notebook step, parameters, AI system and task, human reviewer, output location, and unresolved limitations.

The NIST Research Data Framework organizes research data work across six lifecycle stages, from planning and acquisition through analysis, sharing, reuse, and preservation. Its structure is useful here because AI can affect several stages at once. A model might suggest a source during acquisition, generate code during analysis, and compress results during communication. Each intervention needs its own trace.

The framework's Version 2.0 publication defines provenance as an attributed record of a data asset's origin and alterations, and discusses versioning, parameters, and reproducibility. Those fields belong in the lineage record rather than in a separate retrospective note.

Sources: [1] [2]

Use a four-link reproducibility trail

For each finding, require four connected links:

Do not allow a citation to stand in for these links. A paper can be correctly cited while its sample, variable, or caveat is used incorrectly. Likewise, retaining a chat transcript does not show which input version was analyzed or whether the final claim survived a rerun.

• Source link: the exact document, table, dataset version, inclusion rule, and access condition.
• Transformation link: the code, formula, classification instruction, or manual operation that changed the source material.
• Interpretation link: the claim being made, the supporting output, the analyst's reasoning, and plausible competing explanations.
• Verification link: the independent check, its result, the reviewer, and the remaining uncertainty.

Worked artifact: a synthetic multi-site evidence memo

A fictional research group is comparing water-quality readings from three monitoring sites. An analyst gives an AI assistant two comma-separated files and asks for the largest year-on-year change. The assistant reports that Site C improved by 18 percent and drafts an optimistic paragraph.

The lineage record exposes three problems. The two files use different unit labels, the later file contains a revised batch with the same filename, and the assistant excluded blank readings without recording the rule. The analyst creates finding WQ-017 and records:

The corrected memo reports the observed difference and the sensor change together. The 18 percent figure is not deleted merely because AI produced it; it is withheld because the comparison is not yet supported by a stable measurement basis. This example is synthetic and does not establish that the underlying research design is valid.

• source versions 2025-04-v2 and 2026-04-v3, with checksums and repository links;
• a conversion from micrograms per litre to milligrams per litre for one column;
• the row filter and missing-value rule in notebook cell 14;
• the AI task as code explanation and draft wording, not statistical validation;
• the computed change, confidence interval, and sensitivity result;
• a reviewer note that Site C changed sensor hardware between periods;
• a disposition of revise, because the original sentence implies comparability that the evidence does not establish.

Verify by rerunning and substituting

Use two checks before a finding leaves internal review. First, a rerun check: a second person should be able to recreate the output from the recorded source versions, transformation, parameters, and environment. Second, a substitution check: replace the AI-generated extraction, label, or code with an independent human check or a different method for a risk-weighted sample.

Record agreement, disagreement, and the effect on the conclusion. A disagreement that does not change the decision can remain as a documented limitation. A disagreement that changes the direction or strength of the finding should send it back for revision.

Keep the test proportionate. A low-stakes discovery list may need sampled verification. A central quantitative claim may need complete reconciliation, domain review, and an archived execution environment. The point is not to create paperwork for every prompt. It is to retain enough evidence for the important findings to be challenged and recreated.

Decide what can be released

A finding is ready only when its source and transformation are identifiable, the claim matches the output, material caveats are visible, the verification result is recorded, and access or sharing restrictions are respected. Use clear dispositions such as ready, revise, exclude, or method review required.

AI assistance does not transfer accountability from the researcher or organization. Two Prune can assess how a participant frames a research task, handles evidence, uses AI, verifies outputs, and communicates uncertainty inside a bounded simulation. It does not certify research validity, conduct peer review, or demonstrate that an organization meets a funder's or regulator's requirements.