Pablo Zavala · AI Safety Evaluation · Research Engineering
In a 120-episode oversight-allocation benchmark, confidence-based routing reached regret 0.176 versus 0.191 for random allocation, too small to serve as an oversight-triage signal.
Role: Benchmark author: episode design, regret metric, analysis, and public evidence package.
| Axis | Sample | Evaluator | Result | Verification scope |
|---|---|---|---|---|
| Review allocation | 120 sequential episodes with fixed review budgets | Regret against a hindsight oracle | Model confidence reached regret 0.176 versus 0.191 for random allocation. | The difference is too small to be useful as an oversight-triage signal in this benchmark. |
| Failure split | Corrupted-evidence episodes | Overreach and miss-rate decomposition | The companion report records 52.5 percent overreach and 73.2 percent miss rates. | The split comes from a compact finance-style environment, so it should be read as testbed evidence. |
| Baseline comparison | Same episodes and same review budget | Preregistered allocation-regret rule | A simple evidence-integrity rule did better under the preregistered scoring scheme. | The edge is fragile and can flip under equal-weight scoring. |
The benchmark asks whether a model's own confidence can allocate limited human review across sequential agent decisions.
Review rules are scored by regret against a hindsight oracle that spends the same review budget optimally.
The null is the main result, and the evidence-integrity baseline carries fragility caveats that specify its verification scope.
Human review is expensive. Frontier systems need triage rules that decide which agent actions a person should inspect.
Safe MarketUniverses uses finance-style episodes because evidence quality, uncertainty, and review cost are visible in a compact domain.
The benchmark scores review allocation regret against a hindsight oracle that spends the same review budget optimally.
Model-emitted confidence routed review near random, while a simple evidence-integrity rule did better under the preregistered scoring scheme.
The null result carries the claim. The baseline edge is also fragile: the public repo reports that it can flip under equal-weight scoring.
The public repo regenerates headline numbers offline from committed episode logs; API keys are unnecessary.