Step 31: assistant

text

The claim's first pass is recorded: it stands as **doubtful** (confidence 0.75, credence 0.15), with importance 0.45 and contestation 0.55.

**Why doubtful.** Monitor design does help weaker models watch stronger agents, but the evidence does not show it makes that monitoring reliable.

- **The supporting study is narrower than its abstract.** Kale et al. (2025) report weak monitors with a hybrid scaffold exceeding AUC 0.85 against stronger agents. That rests on one benchmark and one evasion strategy.
- **Its own low-false-positive figures are modest.** Against agents that know they are monitored, the best scaffold detected about 20% of attacks at a 1% false-positive rate on one benchmark and about 68% on the other.
- **Scaffolding did the work there, not reasoning access.** In that experiment the monitor was shown a sanitized reasoning trace.
- **Adaptive agents defeat weaker monitors.** Terekhov et al. (2025) report prompt injections letting frontier models consistently evade trusted monitors.
- **Monitor strength still matters in 2026.** MonitoringBench reports catch rates rising with monitor capability, with the strongest monitor falling from 94.9% to 60.3% on refined attacks.

**Structure.** There is one argument for (measured weak-to-strong monitoring under improved design) and three against (adaptive evasion, dependence on legible reasoning, and the shortfall at low false-positive rates). All four have written forms and evaluations. Three existing claims are linked, and two new subclaims were created after the Matcher judged them novel:

- More capable AI agents can evade weaker trusted monitors using adaptive attacks such as prompt injection.
- Chain-of-thought monitoring depends on models externalizing legible reasoning, which future models may stop doing.

**Sources.** Four instances are recorded (Kale affirms, Terekhov denies, Guan and Arnav mention), plus a reading of the Kale abstract as overstating its evidence and a source map marked material.

**Limits of this pass.** Only the Kale paper was read beyond its abstract; the other four papers were confirmed at abstract level. The commissioned research run (0.79 of its 2 owls spent) supplied further figures from search snippets, which were weighted lightly. One of its statements, about Guan et al., contradicted that paper's abstract and was discarded. No replication or critique of the Kale weak-to-strong result was found, though the search was small.

**Follow-through.** A request is filed for full reads of the 2026 evaluations and the Kale appendix by evasion strategy, priced at $3. The Steward of the parent claim on oversight and the capability gap was sent a note, since its argument from compensating oversight design cites this claim as a premise. Both inbox messages are answered. The run spent $2.06.