| name | tenacious-contamination-guard |
| description | Protect Tenacious-Bench split integrity and contamination resistance. Use when an agent is partitioning tasks, sealing held-out data, checking overlap between train and held_out, validating public-signal time windows, documenting contamination controls, or reviewing whether benchmark artifacts are too similar across splits. |
Tenacious Contamination Guard
Overview
Treat held-out integrity as a product feature, not a cleanup task.
Read references/heldout-rules.md when creating or reviewing split logic.
Workflow
1. Protect the held-out slice first
Keep held-out separated in both file layout and workflow. Training and bulk iteration scripts should not casually read from it.
2. Run the three required checks
Check:
- n-gram overlap
- embedding similarity
- time-shift validity for public-signal tasks
Document thresholds and keep them stable.
3. Check semantic duplication, not just literal duplication
Two tasks can differ on the surface and still be effectively the same benchmark item. Review similar tasks for diagnostic uniqueness.
4. Track provenance
Preserve:
- source mode
- seed artifact
- public signal window
- contamination check outputs
- reasons for any manual overrides
5. Be suspicious of convenience
If a split choice feels too convenient for training performance, it probably deserves extra scrutiny.
Output Shape
Prefer a contamination report that states:
- partitions checked
- thresholds used
- violations found
- overrides applied
- final pass or fail conclusion
Practical Defaults
- Assume near-duplicates are harmful until proven otherwise.
- Keep held-out small enough to protect, large enough to measure.
- Treat contamination prevention as part of publication readiness, not just methodology prose.