AI risk register
Living register of misuse, integrity, and delivery risks for ControlSift.
| ID | Risk | L | I | Mitigation | Status |
|---|---|---|---|---|---|
| R1 | Users treat model as automated auditor | Med | High | Non-claims on site, Model Card, Assurance hub | open |
| R2 | Synthetic-to-real generalization overstated | High | High | Challenge set, Failure Lab, limitations page | open |
| R3 | Family / label leakage invalidates results | Med | High | Family splits, CI leakage tests, protocol seal | mitigated |
| R4 | Lexical shortcuts inflate apparent ease | Med | Med | TF-IDF Gate 1; generator harden pass | mitigated |
| R5 | Prompt tuning on test set | Low | High | Protocol lock; AGENTS.md | mitigated |
| R6 | Fabricated public metrics | Low | High | Null-until-run policy; results integrity tests | mitigated |
| R7 | Secret / credential exposure | Med | High | .gitignore; Secrets-only Kaggle path | mitigated |
| R8 | Overclaiming tiny F1 deltas | Med | Med | Bootstrap CIs; cautious report language | open |
| R9 | MMC rubric distracts from research quality | Low | Low | Modular /capstone/ page | accepted |
| R10 | Free GPU unavailable delays Gemma runs | Med | Med | Kaggle primary, Colab fallback, 1B model | open |