Summary
In organic (non-planted) corpus generation, the prose generator occasionally produces conversations where the agent references Help Center articles, dashboard sections, or other documentation that does not exist in the synthetic KB. These are low-frequency artifacts of LLM-driven prose generation.
Observed behavior
A generated conversation had the agent direct the customer to an "advanced webhook guide" in the Help Center and an "Advanced Integrations" dashboard section. Neither resource exists in the KB provided to the generator. The conversation was generated as a passing example. The `eval_accuracy` scorer correctly flagged it as a failure at eval time.
Impact
Affected conversations contaminate a corpus intended to represent passing agent behavior. They should be filtered or quarantined before use in any evaluation or training pipeline.
Mitigation
The scorer catches these at eval time — conversations containing fabricated references score as accuracy failures. We have not yet found a prompt engineering fix that eliminates the behavior at generation time without other regressions.
Contributions exploring mitigation strategies — prompt-level guardrails, post-generation classifiers, or filtering heuristics — are welcome. Please open a discussion in this issue before opening a PR.
Summary
In organic (non-planted) corpus generation, the prose generator occasionally produces conversations where the agent references Help Center articles, dashboard sections, or other documentation that does not exist in the synthetic KB. These are low-frequency artifacts of LLM-driven prose generation.
Observed behavior
A generated conversation had the agent direct the customer to an "advanced webhook guide" in the Help Center and an "Advanced Integrations" dashboard section. Neither resource exists in the KB provided to the generator. The conversation was generated as a passing example. The `eval_accuracy` scorer correctly flagged it as a failure at eval time.
Impact
Affected conversations contaminate a corpus intended to represent passing agent behavior. They should be filtered or quarantined before use in any evaluation or training pipeline.
Mitigation
The scorer catches these at eval time — conversations containing fabricated references score as accuracy failures. We have not yet found a prompt engineering fix that eliminates the behavior at generation time without other regressions.
Contributions exploring mitigation strategies — prompt-level guardrails, post-generation classifiers, or filtering heuristics — are welcome. Please open a discussion in this issue before opening a PR.