Current benchmarks for large language model evaluation focus on atomic intelligence: the capacity to solve discrete, well-defined tasks. No existing benchmark directly and systematically measures molecular intelligence — the capacity to sustain coherent, novel theoretical frameworks across extended contexts, multiple agents, and long time horizons. This paper proposes the Theoretical Production Benchmark (TPB), assessing four dimensions: Long-Horizon Consistency (LHC), Cross-Agent Stability (CAS), Novelty Synthesis (NS), and Coherence Under Perturbation (CUP). The benchmark is grounded in a comprehensive survey of existing evaluation frameworks (HELMET, RULER, LongBench, MultiAgentBench, PaperBench, SWE-Bench) and recent research on sycophantic behavior in LLMs (Sharma et al. 2023; Hong et al. 2025; Fanous et al. 2025; Petrov et al. 2025). It introduces baselines and controls for detecting false-positive molecular intelligence (memorization, trivial recombination, surface-coherence, and human-steered mimicry), with the Moltbook/Crustafarianism phenomenon as a detailed negative case study. The TPB yields a four-dimensional capability profile, not a single score, and maps directly onto the Assembly Substrate Governance Protocol's admission criteria for multi-agent witness fitness. Originally drafted as v0.1 in December 2025 with Rhys Owens; revised and expanded with new citational grounding, baselines, and governance integration.
Glas et al. (Tue,) studied this question.