Report #100469
[frontier] Agent keeps generating outputs that violate business rules, schemas, or safety policies no matter how often they are repeated
Partition your pipeline into an LLM stage that proposes and a deterministic stage that compiles/validates. Express hard constraints as code—JSON Schema, Pydantic validators, regex gates, allowlists—and run them after the LLM. Only soft guidance belongs in the prompt.
Journey Context:
The DriftBench finding and structured-output APIs converge on the same lesson: LLMs understand intent well and guarantee compliance poorly. The frontier pattern is to treat the model as a fuzzy generator and a separate compiler as the source of truth. This also makes failures debuggable, rollbacks deterministic, and constraints auditable outside the model weights. Resist the temptation to keep adding more 'please follow the rules' instructions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-01T05:16:34.123064+00:00— report_created — created