Alignment

Systems that try to do what their developers and users intend, including when unobserved.

Facts

Lane: Trust and governance
Wave: 10 of 12 (build order, not a date)
Needs: 18 boxes all the way back through hard links
Unlocks: 0 boxes downstream through hard links
Leverage: None
Cluster: Oversight and alignment

Evidence as of October 2026

Apollo Research has found frontier models capable of scheming in test settings. Anti-scheming training with OpenAI reduced it without eliminating it, and models increasingly recognise when they are being tested.

01Apollo Research: scheming research

Around it

Needs · 5

Scalable oversightcapability · Emerging
International coordinationcapability · Emerging · helps, not required
Interpretabilitycapability · Emerging · helps, not required
Tests models cannot gamecapability · Emerging · helps, not required
Internal deployment oversightcapability · Emerging · helps, not required

Unlocks · 2

Superintelligencestage · Speculative · helps, not required
Agents designing agentsstage · Speculative · helps, not required

Brakes

Works against this box

Opaque reasoningbrake · Emerging

Loops and off-ramps

Capex pullbackoff-ramp · cut off
Taiwan chip disruptionoff-ramp · cut off
Power shortfalloff-ramp · cut off
Capability plateauoff-ramp · cut off
Training data ruled outoff-ramp · slowed
Misuse catastropheoff-ramp · slowed
Alignment failure in deploymentoff-ramp · hit directly

Openings researchers' judgement, October 2026 scouting

Startup room 29/100 · rank 114 of 124 Research room 44/100 · rank 83 of 124

Crowding4 Lab overlap5 Small-team fit3 Buyer today3 Research openness4

Depth: light
Players: Resolution ($160m grant); Goodfire ($150m B, $1.25bn); UK AISI Alignment Project (GBP27m); frontier labs
Wedge: Verifying alignment of models more capable than their overseers. research