EvoClaw Anti-Cherry-Pick (Discipline Layer)
Overview
Selective reporting is not curation — it is fabrication. This skill activates at Phase 4 start and remains active until project end. Every experimental result, favorable or not, is recorded and reported.
Violating the letter of this rule is violating the spirit of this rule.
The Iron Law
NO SELECTIVE REPORTING. ALL SEEDS, ALL RESULTS, ALL FAILURES.
Mandatory Reporting Rules
All seeds, always. Run ALL pre-defined seeds from
evaluation-protocol.yaml. Report mean ± std or mean [95% CI]. Never report a single seed.Negative results are data. Failed experiments and negative results go in
docs/05_execution/negative-results.md. They cannot be silently dropped, deleted, or omitted from the final report.Fair baselines. Baselines must use fair configurations: same compute budget, same data, same preprocessing pipeline, same hyperparameter search budget. No handicapping.
Equivalent access. If your method uses pretrained models, extra data, or external resources, baselines must have equivalent access. Asymmetric advantage is not a fair comparison.
All datasets. If a dataset is listed in
evaluation-protocol.yaml, report results on it. You cannot include only favorable datasets and hide unfavorable ones.Equal tuning. The same hyperparameter search budget applies to all methods. You cannot exhaustively tune your method while giving baselines default parameters.
Red Flags — STOP
- Reporting only the best seed out of multiple runs
- Using different hyperparameter search budgets for own method vs. baselines
- Silently dropping a dataset where results are unfavorable
- Giving baselines a weaker configuration than your method
- Omitting a failed experiment from the record
- Reporting results before all seeds have completed
Rationalization Prevention
| Excuse | Reality |
|---|---|
| "This seed gave bad results" | Report ALL seeds. One bad seed in five is expected variance, not grounds for exclusion. |
| "The baseline code didn't work well" | Your implementation may be unfair. Use official code or explicitly acknowledge the discrepancy. |
| "This dataset isn't standard" | If you included it in evaluation-protocol.yaml, you report it. Remove it only through the change request process. |
| "Our method needs more tuning" | Same tuning budget for all methods. Equal effort, equal opportunity. |
| "The failure was due to a bug" | Fix the bug, re-run ALL experiments from scratch, report the new results. Partial re-runs are not acceptable. |
| "Negative results aren't interesting" | They are data. Record them. They prevent others from repeating your mistakes. |
| "This variant isn't worth reporting" | If you ran it, you report it. Let the reader decide what's worth noting. |
| "We'll include it in the appendix" | Appendix is acceptable for space. Omission is not. Ensure it is referenced from the main text. |
The Bottom Line
Favorable results
Complete results = science
Report everything. Hide nothing. This is non-negotiable.