Draft-Verify
An agent architecture pattern that first generates a draft, then performs symbolic verification of the citations in the draft, and carries out revisions.
Timeline (last 90 days)
In the 240-question test, ReAct's accuracy was 59.6%, Plan-execute's 61.7%, Supervisor's 65.0%, and Draft-verify's 63.7%.
ReAct consumed an average of 4,663 tokens per question, Plan-execute 10,535 tokens, Supervisor 7,494 tokens, and Draft-verify 4,905 tokens.
Draft-Verify's low 0.4% fabricated-citation rate is a property of the experimental setup (verification using the same registry as scoring), not an independent achievement of the verification mechanism.
All Facts (3)
In the 240-question test, ReAct's accuracy was 59.6%, Plan-execute's 61.7%, Supervisor's 65.0%, and Draft-verify's 63.7%.
50%UnverifiedReAct consumed an average of 4,663 tokens per question, Plan-execute 10,535 tokens, Supervisor 7,494 tokens, and Draft-verify 4,905 tokens.
50%UnverifiedDraft-Verify's low 0.4% fabricated-citation rate is a property of the experimental setup (verification using the same registry as scoring), not an independent achievement of the verification mechanism.
50%