AI State of the Field — Document Frontier

Origin: research/AI-State-of-the-Field-REP.md

Knowledge and limitations

The REP argues that verified outcomes, context policy, harness effects, trajectory evidence, and security boundaries dominate current AI-ROS decisions. Its evidence is recent and traceable but concentrated in vendor reports, preprints, public benchmarks, and software tasks; transfer to AI-ROS is untested.

Five highest-value opportunities

1. RFR-004: Test public-to-native task transfer (EV-001, EV-002, “Risks and Research Debt”).
2. RFR-003: Validate layered outcome/trajectory grading (EV-005, EV-006).
3. RFR-005: Causally compare context policies (EV-004).
4. RFR-008: Validate least-privilege containment (EV-007).
5. RFR-006: Measure verified cost rather than token cost (“Recommended Next Research”).

Challenge: The REP’s mission selection is reasoned prioritization, not evidence that evaluation work has higher realized ROI. RFR-006 should eventually test that assumption.