Map/Measurement and understanding/Wave 1

Real-world task evals

Measuring models on real deliverables, so delegation rests on evidence.

Facts

Lane: Measurement and understanding
Wave: 1 of 12 (build order, not a date)
Needs: 0 boxes all the way back through hard links
Unlocks: 3 boxes downstream through hard links
Leverage: None
Cluster: Oversight and alignment

Evidence as of October 2026

OpenAI’s GDPval grades models on 1,230 real work products from 44 occupations; the best model matched or beat human experts 47.6% of the time. METR’s time horizons track how long a task an agent can finish.

Detail boxes 1

Benchmarks for multi-agent organisations capability · Emerging · helps, not required

Tests of teams of agents, and mixed human-agent teams, on dividing work, handoffs and error correction over long horizons, rather than solo leaderboards.

TeamBench (MIT/Google, May 2026): verifier agents approved about 49% of work the grader rejected, and the top model passed 37.8% as a team versus 35.6% solo.

Around it

Needs · 0

Nothing beyond the detail list. It can start at any time.

Unlocks · 5

AI R&D capability evalscapability · Emerging
Calibrated abstentioncapability · Emerging
Agent-native organisationsstage · Emerging · helps, not required
AI tutoring at scalestage · Emerging · helps, not required
AI in public servicesstage · Emerging · helps, not required

Loops and off-ramps

Trust and delegationreinforcing loop

Openings researchers' judgement, October 2026 scouting

Startup room 36/100 · rank 86 of 124 Research room 35/100 · rank 106 of 124

Crowding4 Lab overlap4 Small-team fit3 Buyer today4 Research openness3

Depth: light
Players: Mercor ($20bn talks); OpenAI GDPval; Artificial Analysis; Epoch
Wedge: Cheap, reliable grading of open-ended professional work with external validity. research