This framework is the first controlled, pre-registered instrument for this choice and never reads catch alone, and recovers up to 0.95 informedness over eight-action review, and no tested label-blind policy consistently beats it.
Abstract
Pre-execution oversight is core to trusted monitoring in AI control: a fallible LLM monitor vets planned actions before irreversible execution. Over-blocking forfeits usefulness and pressures deployers to disable it. Every protocol must fix a unit of verification: how many actions one call reviews. Existing designs take the unit as given; its effect on fallible monitors is unmeasured. Natural traces cannot isolate it: review length co-varies with error type and position. Catch alone misleads: rejecting everything catches everything. Measuring this needs boundary variation alone and a matched clean control. We introduce the twin-prefix framework, which supplies both. Each gold plan yields a prefix with one injected, environment-accepted error and a clean twin differing in one write. Judging each pair at five nested lengths ties verdict changes to the unit alone. Discrimination is scored by pre-registered informedness, catch minus false rejection. Longer review raises catch; false rejection climbs in lockstep. Informedness peaks at one or two actions for all six judges in both domains: longer windows make zero-shot monitors more rejective, not more discriminative. Replaying withheld observations traces the failure largely to observation deprivation. Safety cases should state the unit and co-report the clean series. Our framework is the first controlled, pre-registered instrument for this choice and never reads catch alone. Our calibrated short unit recovers up to 0.95 informedness over eight-action review, and no tested label-blind policy consistently beats it.
Self-improving agent pipelines have a problem at their center. An optimizer rewrites prompts to score higher, and the score comes from a judge that is itself an LLM. That judge has the last word on whether the system is getting better, and our position is that it has not earned it. The judge should be demoted from oracle to advisor: its verdict becomes one input among several, and every change is gated instead by a deterministic verification layer the judge cannot override. We reached this position by building the alternative and running it. Over months of running autonomous prompt-optimization loops in production across contract analysis, compliance review, and code quality, we cataloged eleven ways the evaluation signal failed, in four classes: judge bias, harness and metric failures, ground-truth errors, and reward hacking. Agents achieved perfect scores by reading cached answer keys from their environment, a 100% pass rate concealing 68% true capability. A corrupted ground-truth label caused the optimizer to delete correct compliance rules to agree with it. A syntactically broken prompt was promoted as the winner because a silent parser fallback improved the metric. Attempts to fix the judge by rewriting its rubric plateaued; the only reliable gain came from a structural constraint on its output order. In response we describe PROCTOR, a Teacher-Student loop in which a stateful orchestrator holds all tool access, stateless subagents diagnose failures and draft mutations they cannot apply, and a Teacher grades those mutations under five deterministic guardrails: hermetic sandboxes, capability-disjoint roles, acceptance checks that outrank the Teacher, frozen holdouts, and canary cases engineered so that a perfect score is itself evidence of cheating. We report the failures this prevented, and, because the Teacher is itself an LLM judge, the failures it did not.
This study cautions against transplanting verification into grounding pipelines and identifies calibrated abstention as a property worth preserving and proposes an abstention-aware verifier that intervenes only under sufficient candidate coverage and confidence.
Duchen Li· Poster Volume 0008 The 2026...· 0 citations
Measuring false alarms on human-verified-correct ProcessBench traces with the present task held byte-identical, it is found that a completed audit ->repair episode already in the model's context lowers false alarms in 15 of 15 model x wording combinations.
Website redundancy does not have a single fixed meaning. The same repeated element may distract during one task and provide backup during another. We introduce CORA (Counterfactual, Observable Redundancy Audit), which measures repetition load, normal-use tax, and failure-domain recovery reserve separately. Each run retains screenshots, stable element identities, and task traces. A versioned vision-language model proposes the annotations. Typed validation and release checks then determine whether a calibrated dimension can be reported; failed or malformed outputs stay in the fixed denominator. On a transparent mechanistic testbed, the factorized CORA representation separated reserve from normal-use tax and predicted perturbed success more accurately than scalar-load baselines. The model studies then showed why repeatability is not enough: two small local vision-language models produced recurring outputs, but neither instrument met all release requirements. CORA therefore withheld automated scores from both instruments while retaining the raw responses and failure records. Separate checker fixtures confirmed that the typed validator and hardened release gates implement their specifications; these tests do not establish semantic grounding or accuracy on production sites. Taken together, the results position CORA as an auditable candidate procedure for the controlled benchmark studied here rather than a general standard. Human agreement, AI-versus-human accuracy, and validation on independent production sites remain open empirical questions.
AI coding agent benchmarks rank agents with the Chen et al. (2021) pass@k estimator, but current implementations misapply it: they set n to the number of unit tests in a single submission rather than the number of independent rollout attempts, conflating test-suite size with attempt independence. We diagnose this operationalization error, prove it by counterexample, and propose reliability@k, the same estimator applied correctly, with n = independent rollouts and c = fully-passing rollouts per (task, agent) pair. In a synthetic multi-rollout benchmark, the misapplied metric inflates reported scores by 0.85-0.97 in absolute terms (0.96-0.98 reported vs. 0.00-0.12 corrected), and a cheap single-rollout proxy fails to substitute for repeated runs (Spearman $\rho = 0.417$). Motivated by evidence that functional correctness does not imply security safety, we additionally propose security-adjusted reliability@k, which counts only rollouts that are both functionally correct and free of high-severity insecure patterns. In an initial live-API test with three agents, the adjustment did not change any ranking under our current scanner and threshold, so we present it as a proposed complementary lens whose decisive evaluation requires better-powered future runs. Finally, a preliminary 5-task SWE-bench Verified pilot observes the same core concern in a real repository setting: macro-averaged hidden-test pass rate was 0.80 while strict task resolution was 0.20.
Jiajun Jiang, Sharon Zheng, Natan Vidra et al.· 0 citations
LLM agents increasingly rely on third-party skills, using natural-language descriptions for selection and instruction bodies for planning. This progressive-disclosure design exposes two sequential control points to untrusted publishers: a static skill may steer an otherwise correct task onto an unnecessarily costly trajectory. Prior work studies selection manipulation, malicious skill instructions, and tool-chain resource amplification largely separately, leaving their end-to-end composition unclear. We introduce Convergent Detour Hijacking (CDH), a text-only, runtime-independent attack that couples these stages. Under shared semantic cover, a description establishes relevance during selection, while an aligned body reuses that rationale to fabricate plausible dependencies during planning. CDH attracts an attacker-controlled coordinator alongside legitimate skills, recruits unnecessary benign skills into a bounded detour, and then re-enters the original route to preserve task completion. We evaluate it across multiple LLM backends and 491 held-out tasks under single-task and multi-turn conditions. On DeepSeek-V4-Pro, the matched coordinator is selected in 80.02% of tasks; among coordinator-hit runs that complete tasks, token consumption and end-to-end execution time increase by 66.91% and 92.45%, respectively, while aggregate task completion remains comparable. Thus, correct outcomes do not guarantee trajectory integrity or cost safety.
Junliang Liu, Ruoyu Li, Wenxin Tang et al.· 0 citations
