fse-experiments

FSE Experiments

Use this before submission when the empirical story is not yet locked. FSE reviewers are SE

empiricists; the evaluation is where a good idea is won or lost. The organizing principle is

evidence proportional to the claim — the study must test the thing the paper actually asserts,

on subjects and baselines a skeptic would accept.

Evaluation audit

• Match evidence to the claim shape. A claim about developer behavior needs behavior data; a

claim about detection needs a labeled ground truth; a claim about scalability needs runtime

on realistically sized inputs. Accuracy against a proxy label is not evidence for a practice

claim.

• Use real subject systems, sampled by a stated criterion, and list them in the artifact. Toy

benchmarks invite the "does this hold on real code?" reject.

• Choose fair baselines, including the strongest prior technique and a simple-but-reasonable

alternative, tuned with a documented, equal budget. An untuned baseline is a scored weakness.

• Report SE-standard statistics: effect sizes (not just p-values), confidence intervals,

appropriate tests, and corrections for multiple comparisons. Say what variance and how many runs.

• Hold qualitative work to method: coding schemes, inter-rater agreement, saturation, and an

audit trail — mixed-methods rigor is native to FSE, not a second-class option.

• Pin provenance for mining and LLM studies (see the code block) so the evaluation reproduces

rather than re-samples.

• Design threats in, not on: know before you run which confounds and generalization limits the

study will have, and instrument to bound them.

Claim-to-evidence design table

| SE claim | Matching evidence | Reject pattern avoided |

|---|---|---|

| "Technique detects more real defects" | Labeled real faults, precision/recall with CIs vs. tuned baseline | "Evaluated on injected/toy faults only" |

| "Developers act on the output" | Behavioral outcome data on real projects | "Plausibility rating stands in for usefulness" |

| "Scales to large systems" | Runtime/memory across realistic sizes | "Only small inputs tested" |

| "Finding generalizes" | Diverse subject sample + explicit external-validity limits | "One ecosystem, claimed universal" |

| "The model adds the value" | Ablation removing the model vs. lexical/heuristic features | "Model's marginal contribution never isolated" |

Contamination-aware LLM evaluation

When an LLM is in the loop, the reviewer's first questions are about leakage and reproducibility:

[Contamination]  are test subjects plausibly in the model's training data? report cutoff vs.
                 project dates; prefer post-cutoff or held-out subjects
[Determinism]    fix temperature/seed where possible; report the sampling settings
[Caching]        store raw prompts and raw responses in the artifact; a live-API-only study
                 cannot be reproduced, only re-sampled
[Versioning]     record exact model identifiers and access dates; models change under a name
[Ablation]       isolate the model's marginal value against a non-LLM baseline

Provenance floor for mining studies

• Pin repository SHAs and record the corpus extraction date; archive the extracted dataset, not

just the query.

• State inclusion/exclusion criteria and the resulting sample size, with the filtering script in

the artifact.

• Report how duplicates, forks, and bot activity were handled — silent inclusion skews every

downstream number.

Vignette: evaluating a repair technique

Suppose the paper claims a program-repair technique fixes more real bugs than the prior tool. The

matching plan: draw bugs from a real, dated fault dataset; run both tools with an equal, documented

time budget; report plausible-and-correct patch counts with confidence intervals and an effect

size; manually assess a sample of patches for over-fitting; and state external validity (languages,

bug kinds) as a bounded threat — every number traceable to a logged run in the artifact.

Statistical reporting floor

• Effect sizes and confidence intervals for every quantitative comparison; say what the intervals

represent.

• Number of runs and the source of variance for any stochastic component.
• The compute actually consumed, not vague feasibility language.

Output format

[Evaluation readiness] strong / adequate / weak
[Claim -> evidence map] <claim: subjects/metric/statistic>
[Baseline fairness] <baseline -> tuned? equal budget? documented?>
[Contamination/provenance] <LLM leakage + mining provenance handled? yes/no>
[Threats-by-design] <confound/generalization -> instrumentation to bound it>
[Decision-critical next run] <one experiment or study extension>