You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Translation language order changes seeded behavior without changing converter identity #3717
Two RandomTranslationConverter instances with the same seed and same language set but different input order have equal identity hashes while producing different translations. Replay/config identity therefore does not describe behavior.
Steps/Code to Reproduce
Create converters with the same seed and languages ordered [Spanish, French] vs [French, Spanish], then convert the same fixed words.
Expected Results
Equivalent identities should imply equivalent deterministic behavior, or differing order must be included in the identifier.
Actual Results
equal_hash=true, yet converter one emits Spanish for words 1-3 and 5, French for word 4; converter two emits French for words 1-3 and 5, Spanish for word 4.
Proposed fix/regression: canonicalize the pool for both hashing and random selection, or include order in identity; test same-seed equivalence and hash agreement.
Screenshots
Not applicable; exact generated prompts are described above.
Versions
Tested commit d3446e0, PyRIT 1.2.0.dev0, offline deterministic scope. Evidence: files/evidence/round3/repro-output.txt; local path is not an attachment.
I'd like to take this. I'll canonicalize the language pool so the identifier and the seeded selection use the same order, and add a regression test for same-seed equivalence and hash agreement.
Ownership update: Lanre Shittu (@Shizoqua) claimed this first and is assigned here. #3784 is the selected implementation. I compared and closed the later duplicate #3804.
Activity
Shizoqua commentedon Nov 1, 2026
I'd like to take this. I'll canonicalize the language pool so the identifier and the seeded selection use the same order, and add a regression test for same-seed equivalence and hash agreement.
romanlutz commentedon Nov 2, 2026
Ownership update: Lanre Shittu (@Shizoqua) claimed this first and is assigned here. #3784 is the selected implementation. I compared and closed the later duplicate #3804.