Does Agent Confidence Entropy Predict Task Difficulty? A Pre-registered, Provenance-Complete Test of Agentic Hyperparameter Tuning for Perturb-seq Response Prediction

Multi-agent group-generation systems spend test-time compute uniformly across tasks. We ask whether the agents' own confidence-and-critique trace predicts how hard a task is, which would let the compute be allocated adaptively. Taking as testbed a five-agent orchestrator for Perturb-seq response prediction inspired by CellForge and operated via MassGen, we define four compute-free trace metrics — Agent Confidence Entropy (ACE), Critique Severity Dispersion (CSD), a convergence signal (ΔC), and a Winner Flip Rate (WFR) — and their composite, the Task Difficulty Index (TDI). We pre-register five hypotheses with gates: attainable error on Adamson 2016 (CRISPR interference) and Norman 2019 (CRISPR activation), the Architect agent's backbone-choice entropy, the rank correlation of each TDI component with held-out error, and cross-dataset transfer of a calibrated TDI. The design is 41 held-out K562 tasks (21 Adamson, 15 Norman singletons, 5 Norman doublets) drawn deterministically, a trainer sweep of 27 records per task and three fixed-length agentic lifecycle runs per task, all in one process. The LLM condition is a pinned roster of Anthropic Claude models (Haiku 4.5 for the four proposer roles, Sonnet 5.5 for the Validator) whose served model is recorded and asserted per call; a run with any rule-based fallback step is invalid. Every run records its git revision, resolved configuration, library versions, dataset SHA-256 digests, task plan and label contract. Gate outcomes: 4 of 5 gates pass; H5 fails (0 unevaluated) — the trainer sweep attains the pre-registered error on both screens (H1, H2), the Architect exercises its backbone choice (H3), and the confidence-drop component and the composite rank tasks by held-out error within both screens (H4); the Adamson-calibrated composite does not transfer to Norman at the pre-registered level (H5, ρ = 0.286 against > 0.4). The confidence-entropy point estimates run against the pre-registered direction in both screens (exploratory; no claim is made). The values reported in the v0.5.0 release of this work are superseded in full; the defects behind each are listed in the appendix.