ResearchBenchmarksSteering Validation on Synthetic Stubs Comparison Phase 1Follow96.7Haiku Judge AccuracyDACS36.3852.0467.783.36Apr 9, 2026Evaluation ResultsMethodMethodLinksHaiku Judge AccuracyContamination RateAverage Context TokensP-valueDACSN=3, Scenario=s1_n3N=3, Scenario=s1_n32026.0496.705610.0001DACSN=5, Scenario=s2_n5N=5, Scenario=s2_n52026.0496.706330.0001BaseN=3, Scenario=s1_n3N=3, Scenario=s1_n32026.0460—1,1910.0001BaseN=5, Scenario=s2_n5N=5, Scenario=s2_n52026.0438.7—1,7200.0001