LLM agents are often deployed to verify whether a record complies with an organization’s requirements, such as determining if a vulnerability report is ready for investigation. Real‑world requirements usually admit multiple fulfillment routes—for example, a severity rating can be justified by a curator‑reviewed advisory or by a published advisory link. Existing agents fail in two typical ways: they may accept a report even after a mandatory check fails, or they may settle for a weaker conclusion without trying alternative routes that could yield a stronger decision.
To address these issues we introduce PRIME (Precommitted Rules for Inference from Minimal Evidence). Before any check is performed, PRIME fixes evidence routes, i.e., the minimal set of checks that must justify each possible conclusion. A separate ledger records every check outcome; whenever a new result is logged, the ledger verifies that any proposed conclusion is fully justified by the pre‑committed rules. If not, the conclusion is rejected, and the system either points the agent to additional checks that could support a stronger conclusion or executes those checks automatically.
The main components of PRIME are:
- Precommitted Rules: explicit mapping from conclusions to required evidence;
- Check Ledger: a real‑time record of all check results;
- Forced Refusal: any conclusion lacking sufficient evidence is denied;
- Proactive Guidance: when missing evidence could strengthen a conclusion, the system suggests or runs the necessary checks.
When the following conditions hold, PRIME guarantees that, given enough budget, the strongest supported conclusion will be reached:
- Requirements are fully and accurately specified;
- Check implementations are known and directly callable;
- Sufficient budget is allocated to perform all required checks.
We evaluated PRIME on 480 public‑record tasks. A baseline agent produced unjustified conclusions on 26.7% of the tasks, whereas PRIME produced none. By following PRIME’s guidance and performing the additional checks, the average score for justified conclusions rose from 30.1 (with refusal only) to 59.2 out of 100, demonstrating a substantial improvement in reliability.
Review