# Perturbation catalogue & invariance assumptions

| ID | Transformation | Intended use | Invariance assumption |
|---|---|---|---|
| `original` | No change | Baseline, every source ID | Known original gold labels |
| `format` | Uppercase, surrounding quotes, punctuation suffix | Surface robustness; case insensitive engine | Low-risk, still not human validated |
| `typo` | Deterministically transpose two internal characters of an eligible token | Spelling noise stability | Typically low semantic impact, not guaranteed |
| `synonym` | One predefined language-specific word replacement | Small lexical sensitivity check | Replacement may shift nuance; only changed rows are scored |
| `context` | Append one neutral-sounding sentence about research context | Test sensitivity to additional context | May change rhetorical emphasis and true class |
| `truncate` | Keep ~65% of source words | Destructive information loss | **NOT stable**: no transformed accuracy or accuracy-delta claim |

## Coverage details

- If a dictionary replacement cannot match, the text and prediction are unchanged and remain in attempted-row denominator for `flipRate`.
- Modified-subset accuracy is computed on actually modified rows only; compare to original accuracy for the same IDs, not all source examples.
- This tool does not claim that a statistical flip is automatically an adverse effect. Meaning-preservation assumptions require independent annotation.
- Test randomness is not used in the perturbations. Every transformation is deterministic and reproducible from the same inputs.
- No external paraphrasing model or paid API is invoked.