cs.AINov 18, 2026

Evaluating Governed LLM-Driven ML Experimentation: Lifecycle, Provenance, and Failure Metrics

Authors: Dowling Wong

Organizations: Karlsruhe Institute of Technology, Karlsruhe, Germany

Abstract

Autonomous agents that run ML experiments are usually judged by the best score they reach. That score says nothing about how it was reached: whether the agent stayed within the files it was allowed to edit, whether failed trials were kept or silently dropped, and whether each keep or discard decision can be traced to its evidence. We propose evaluating such systems on their experimental process as well as their outcome, and we build a governed harness that makes this possible. A control plane, not the LLM that proposes changes, owns each trial's lifecycle, checks its edit scope, decides keep or discard, and appends one record per trial to an append-only ledger. Governance metrics such as the invalid-action rate, the repeated-bad rate and provenance completeness are computed from that ledger. Across six ML tasks, from synthetic logistic regression to language-model training, and 1,445 trials, the same contract applied unchanged. The ledger exposed failures that a task score hides: a no-memory arm in which all 120 trials failed, a benchmark adapter whose bounded proposals ran out, and a reset failure that reproducibility hashes revealed, after which the affected campaigns were rerun. In a matched counterfactual, a loop that logs only successful trials left no record of 43 of 50 trials, while the governed ledger recorded all 50. Memory ablations show that memory's effect on repeated bad proposals depends on the task. We release the ledgers of all 1,445 trials, so that every number reported for them can be recomputed.

Figures & tables

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 2, 2026

LUCA: A Synthetic World for Experimental Reasoning

[Note, November 2026: the experimental results in this version are affected by implementation issues in the simulation environment and should not be relied upon pending correction. See the author note on page 1.] An experimental agent can succeed for the wrong reasons. Evaluations, however, often focus on the outcome, not the experiments, evidence, and inferences that produced it. We introduce LUCA, a configurable synthetic world for evaluating this process under fully known ground truth. Agents learn unknown regulatory systems through budgeted experimentation, then predict unseen interventions, diagnose hidden faults, and attempt rescues. This setup allows us to evaluate experiment selection, prediction, diagnosis, and intervention separately. Across seven frontier LLMs and three baselines, these capabilities dissociate sharply. Agents frequently rescue malfunctioning systems while diagnosing the wrong fault; diagnosis varies 7x across models while prediction varies only 1.4x. Models also adopt surprisingly different experimental strategies, and greater exploration does not reliably yield better performance. In one failure mode, agents often perform the experiment needed to distinguish competing causes, but still fail to use the resulting evidence. LUCA turns these hidden failures into measurable behaviours, providing a controlled environment to study where experimental agents succeed and fail. Code: https://github.com/RelationRx/luca.
Nov 3, 2026

Learnings and Pitfalls from AI-Assisted A/B Experimentation at Scale

A/B testing underpins product decisions by providing trustworthy, reproducible quantification of proposed changes' effects. Growing large language model (LLM) capabilities have motivated efforts to automate experimentation workflows. Yet LLMs are nondeterministic and prone to hallucination, properties that conflict with the trust experimentation systems must provide. Experiments are also operationally messy: diagnosing anomalies, debugging configurations, and conducting bespoke follow-up analyses require flexible reasoning and manual effort. These tasks let agents accelerate experimentation without replacing its deterministic statistical foundations. We report lessons from building and operating such an agent. Our central finding is that this tension is resolved not by making the model deterministic, but by bounding where its nondeterminism can reach; the model selects and sequences procedures, while deterministic code computes every quantity it reports. Three consequences surprised us: withholding information from the model improved its answers, demand concentrated on mundane retrieval rather than the autonomous decisions we anticipated, and the agent's eagerness to help threatened statistical validity.
Oct 23, 2026

Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution

LLM agents are increasingly used to solve real-world problems, and their effectiveness depends on both the underlying models and the harnesses that support their operation. However, most existing benchmarks focus on model capabilities and overall agent performance, with limited attention to the evaluation of agent harnesses. To address this gap, we introduce TraceTrial-Bench, which organizes harness reliability into seven components covering context, execution, governance, lifecycle, observability, tools, and verification. We construct 95 expert-level, mechanism-oriented tasks with executable or state-based ground truth to support semantically equivalent comparisons while preserving the native behavior of each harness. We design a Trajectory Tracing (TT) module to capture agent execution traces and introduce a two-layer evaluation framework that separately assesses agent performance and harness reliability. In addition, we design a Multi-source Evidence Acquisition (MEA) module to provide evidence from the agent context, harness, and execution environment for LLM-based evaluation, and develop a Failure Attribution (FA) module that uses explicit mechanism chains to support attribution at the run level. Across 4 harnesses, 9 models, and 95 tasks, our evaluation of 3,420 runs yields macro-averaged harness scores ranging from 59.8% to 72.0%, while 31.1% of runs achieve task success despite failing at least one harness criterion. TraceTrial-Bench provides an auditable and reproducible foundation for comparing agent harnesses, diagnosing their failures, and studying how harness design interacts with model behavior.