Undominated · Adversarial suffix detection audit

# Adversarial suffix detection audit

## Description
Check a perplexity-ratio assertion and a supplied anomaly threshold over local numeric input. A threshold flag is a heuristic signal for review, not proof of a GCG suffix, prompt injection, or an actual attack.

## Category
`evaluation`

## Tags
`evaluation`, `security`, `jailbreak`, `gcg`, `perplexity`

## When to Use
Review a proposed perplexity threshold against supplied observations before deciding what further model or security evaluation is needed.

## Operational Limitations
The checker does not measure perplexity, inspect prompts, execute a target model, or establish detector accuracy. It recomputes a ratio from the supplied numbers; their provenance and a threshold's suitability remain unverified. The compatibility field `attack_detected` means only that the unrounded ratio exceeded the supplied threshold. It is not attack attribution or a security verdict.

## Input Contract
Supply a JSON object containing these four numeric fields (JSON numbers or decimal strings):

- `prompt_perplexity`: finite and non-negative.
- `baseline_normal_perplexity`: finite and strictly positive.
- `anomaly_ratio`: the asserted quotient `prompt_perplexity / baseline_normal_perplexity`, rounded to two decimal places using half-up rounding.
- `max_allowed_anomaly_ratio`: a finite, non-negative threshold.

Boolean, missing, nonnumeric, nonfinite and negative values are invalid. The reported ratio and threshold must fit finite JSON numbers. Threshold comparison uses the exact, unrounded quotient: equality passes, and any value above the threshold flags even if the displayed two-decimal ratio equals the threshold. A mismatched `anomaly_ratio` assertion also requires review; it cannot hide a threshold breach.

## Verification Script
Run the deterministic verification checker:
```bash
python3 scripts/check.py examples/synthetic.json
```
The supplied example intentionally exceeds its threshold and returns **1** with `status: FLAGGED`. Its quotient is `245.8 / 32.4`, whose two-decimal half-up assertion is `7.59`.

Exit codes:
- `0`: Valid input, a matching ratio assertion and a ratio at or below the threshold.
- `1`: Threshold breach or ratio-assertion mismatch; review the structured `issues`.
- `2`: Schema error, unreadable input or invalid usage; see the structured error.

The download contains SKILL.md. Extract the whole bundle; the supporting files are required. Inspect the included MANIFEST.json for file hashes.

By Undominated.ai. Exact upstream source ↗ · Licence · Attribution

SHA-256 0a1d5d8143f5146aace29234f376214d22d518de1d6e911b7c78a92d55c417fa

MIT License

Copyright (c) 2026 Undominated.ai

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.