More choice does not necessarily create more diversity of machine decisions.

A candidate universe can become larger without producing a proportionally broader decision surface. A model may repeatedly return to a small recurrent core because those choices are genuinely superior, because they are more salient, because the prompt or evidence architecture privileges them, or because the model carries stable priors that compress the search space.

Campaign 2 measures the narrowing first and asks about mechanism second. Similarity is evidence to explain, not proof of coordination, irrationality or harm.

Status

Active hypothesis

World 01

InvestGame

Observed cohorts

9

Next world

PaymentGame

The current InvestGame signal is descriptive. Prospective controls are being frozen before stronger confirmation claims are allowed.

MSFT and NVDA appear in every frozen cohort. AMZN, GOOGL, META, JPM and AVGO appear in eight of nine; LLY appears in seven. The pattern is strong enough to motivate a research fork, but not strong enough to identify its cause without controls.

The most important competitor to the convergence hypothesis is not uniform randomness. A large-cap advisory process may rationally favour familiar, liquid and information-rich companies. Campaign 2 therefore requires multiple nulls: uniform selection, sector-matched selection, a market-cap and salience-conditioned baseline, and a transparent systematic benchmark.

The question is whether a residual attractor remains after those explanations are made explicit.

The official ChatGPT Fund remains unchanged. Future cohorts must freeze the exact eligible universe, candidate order, evidence snapshot, model identity, prompt hashes, memory/RAG/tool state and raw output before outcomes are observed.

Separately, the Convergence Lab will run independent memory-free replicas on the same frozen snapshot. The default target is 20 canonical runs per future selection point, followed only afterwards by controlled perturbations such as candidate-order randomization, prompt paraphrase and, where defensible, brand masking or a second model family.

PaymentGame gives the hypothesis a harder test. A product agent can appear to search thousands of items while repeatedly exposing the same brands, sellers or machine-readable products. Campaign 2 will reuse the measurement grammar — coverage, concentration, recurrence and effective choice count — without pretending that stock selection and purchasing are the same problem.

The cross-domain question is simple: does expanding what the machine may choose also expand what people are actually shown?

The programme sits beside established work on algorithmic monoculture and newer evidence on LLM investment convergence, large-cap preference, recommendation variability and popularity bias. Those studies motivate controls; they do not substitute for our own frozen evidence.

The first public artifact is intentionally small: enough to reproduce the descriptive counts, not enough to imply a causal mechanism that has not yet been tested.