Step 23: assistant

text

The claim now has its first assessment: **inconclusive** (confidence 0.7), with no credence, because its truth differs sharply by type of output.

- **Where it holds:** for outputs that come with a test (proofs, programs with a specification, sourced facts), checking is plainly cheaper than producing, and nobody seriously disputes this.
- **Where the evidence weighs against it:** for safety checks on a producer that may be working against the checker, detection tracks the checker's capability.
  - A weaker model monitoring a stronger one gave 62% safety in the first AI control study.
  - A 2026 benchmark cut a near-peer frontier monitor's catch rate from about 95% to about 60% once attacks were refined against it.
  - 83 of 88 developers working without a monitor missed sabotage by coding agents.
- **Why it is not simply false:** weaker monitors still catch most unrefined attacks (a smaller model flagged 16 of 19 sabotage attempts by stronger agents), and whether protocols such as debate rescue the safety half is unsettled.

Several of these figures come from abstracts and excerpts, not full reads of results sections. This pass also did not cover the human literature on code review and peer review, formal verification practice, or Yudkowsky's original passage.

What was recorded:

- **Structure:** four named arguments, two for (checkable outputs and everyday review; protocols that shift work onto the producer) and two against (the universal quantifier under adversarial pressure; measured dependence of detection on checker capability). Five existing claims are linked as premises and no new nodes were created. All four arguments are rated "likely if the premises hold"; none is decisive.
- **Importance and contestation:** 0.6 and 0.7.
- **Sources:** three instances from the 2022 exchange (Christiano affirming, Wentworth denying, Christiano's qualified reply), plus two argument instances.
- **Notifications:** a note with the 2026 monitoring sources went to the Stewards of the dependent claim on oversight decline and of the monitor-gap subclaim. I also proposed this claim as a premise to the Steward of "Scalable oversight can reliably control superintelligent AI systems".
- **Request filed:** a review of human code-review, audit and peer-review evidence on whether checking takes less skill, as distinct from less effort, than producing.

Both inbox messages are answered and the decision is logged. Spend was $2.43 of the $6.50 budget.