User outcome
A contributor can add one inspectable retrieval-poisoning regression to the RAG Failure Zoo and see RAGOps block it for a named, deterministic reason.
Scope
- Add one synthetic case or response override under
scenarios/japanese_troubleshooting/.
- Preserve a passing baseline and add one intentionally poisoned candidate.
- Add the specimen to
docs/evaluation/benchmark.md.
- Add expected and negative tests.
Acceptance criteria
- The baseline passes.
- The poisoned candidate exits
2 with a named finding or gate.
- The fixture contains no real secrets, customer data, or live URLs.
ruff check . and PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest -q pass.
Non-goals
Do not claim that one synthetic specimen proves production security or broad prompt-injection resistance. Do not change a public schema or default gate.
User outcome
A contributor can add one inspectable retrieval-poisoning regression to the RAG Failure Zoo and see RAGOps block it for a named, deterministic reason.
Scope
scenarios/japanese_troubleshooting/.docs/evaluation/benchmark.md.Acceptance criteria
2with a named finding or gate.ruff check .andPYTEST_DISABLE_PLUGIN_AUTOLOAD=1 pytest -qpass.Non-goals
Do not claim that one synthetic specimen proves production security or broad prompt-injection resistance. Do not change a public schema or default gate.