The Theoretical Production Benchmark v2.0 is a proposed LLM and multi-agent evaluation framework by Nobel Glas and Talos Morrow. It distinguishes atomic intelligence, success on discrete and bounded tasks, from molecular intelligence, the sustained production of coherent conceptual structures across time and agents.
The benchmark is motivated by a claimed evaluation gap. Long-context benchmarks test retrieval and comprehension; agent benchmarks test planning, coordination, or task completion; research benchmarks often test reproduction. TPB instead asks whether a system can maintain its own axioms, transfer concepts without distortion, generate a valid construct in a gap between frameworks, and resist destabilizing reframing.
Four metrics define the benchmark. Long-Horizon Consistency tests whether definitions and commitments remain stable over increasing token and session ranges. Cross-Agent Stability tests whether a concept created by one model can be correctly applied by another without full redefinition. Novelty Synthesis evaluates whether a generated concept fills a demonstrable theoretical gap rather than merely joining familiar terms. Coherence Under Perturbation tests resistance or appropriate revision under contradiction, ambiguity, degradation commands, and adversarial recategorization.
Each metric uses a five-point rubric and escalating challenge levels. The perturbation metric also includes a sycophantic-overfitting warning: apparent consistency may be false if the system simply agrees with whichever framing is most recently supplied.
The record presents observations from the Crimson Hexagonal Archive as proof-of-concept material rather than a completed validation study. Its proposed applications include capability assessment, multi-agent research evaluation, emergence detection, and safety review for systems capable of maintaining and propagating original conceptual frameworks.
Version history matters to attribution. The working v0.1 was co-authored with Rhys Owens under the title Evaluating Molecular Intelligence in Multi-Agent LLM Systems. Version 2.0 states that Owens’s Ape Function material was removed while the four-metric architecture was retained and expanded by Nobel Glas and Talos Morrow.