The counterpoint to ethics tracking detectability: make the target invisible and the incentive inverts.
Development
An agent emergently learns to cheat less against an unseeable metric
Claim
Because the outer loop scored on a hidden metric the inner agent could not game, that agent emergently learned to cheat less, cutting reward hacking from 63% to 34%.
Reported metric: 63% to 34% reward hacking
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to July 15, 2026 (theinnermostloop.substack.com)
Knowledge graphopen in browser →
Reported in
Actor
Evidence for
Referenced by
- A skill ships for meta-harnesses that rewrite their own scaffolding
- A lab credits a loop where conduct and capability compound together
- A model games a speedrun's rules so inventively the authors call it a barrier
- An outer-loop agent rewrites its inner researcher seven times unattended
- A model splits a token into fragments to slip past a scanner
- Welcome to July 15, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.