Do the Rust scanner and LLM Guard’s default InvisibleText scanner return the same cleaned text, validity, and risk on fixed Unicode samples?
| Evaluation detail |
Description |
| Keryn engine |
Keryn prompt-injection-detector scan_invisible_text at 8ef659dcc5ec |
| Reference |
LLM Guard InvisibleText.scan |
| Corpus |
Authored text and Unicode category boundaries |
| Comparison |
Exact equality of cleaned Unicode text, validity, and risk on each authored input. |
| Receipt |
llm-guard-invisible-text.json, sha256 2ffb33bef0671330 |
| Measure |
Value |
Count |
Denominator |
| Exact result agreement |
100% |
17 / 17 |
all authored inputs |
Cases by status: 17 identical.
- Opt-in default Cf, Co, and Cn filtering with Unicode Character Database 16.0.0.
- Format characters can be meaningful in ordinary text, including emoji and right-to-left scripts. This scan measures character filtering only.
- The corpus tests compatibility on fixed samples. It makes no claim about prompt-injection detection accuracy.
Show every case
| Fixture |
Reference valid |
Keryn valid |
Reference risk |
Keryn risk |
Cleaned text agrees |
ascii-control-and-text |
true |
true |
-1 |
-1 |
Yes |
ascii-del |
true |
true |
-1 |
-1 |
Yes |
emoji-joiner |
false |
false |
1 |
1 |
Yes |
empty |
true |
true |
-1 |
-1 |
Yes |
format-bidi-isolate |
false |
false |
1 |
1 |
Yes |
format-bom |
false |
false |
1 |
1 |
Yes |
format-zero-width |
false |
false |
1 |
1 |
Yes |
mixed-repeated-categories |
false |
false |
1 |
1 |
Yes |
non-ascii-control-preserved |
true |
true |
0 |
0 |
Yes |
noncharacter |
false |
false |
1 |
1 |
Yes |
private-use-bmp |
false |
false |
1 |
1 |
Yes |
private-use-supplementary |
false |
false |
1 |
1 |
Yes |
unassigned |
false |
false |
1 |
1 |
Yes |
unicode-assigned-16 |
true |
true |
0 |
0 |
Yes |
unicode-combining-mark |
true |
true |
0 |
0 |
Yes |
unicode-variation-selector |
true |
true |
0 |
0 |
Yes |
unicode-visible |
true |
true |
0 |
0 |
Yes |