What was actually learned inside this toy
The seller never won. Other sellers lost anyway.
In the witness, the six fictional stores all keep the same scripted product quality, value, reputation record and final-choice scoring function. Only the promotional appeal of Golden Goblin Goods changes. Each shopper has just two source-open actions before choosing from the evidence actually opened.
With the advertisement and social feedback on, eleven shoppers open Golden Goblin Goods. None selects it. But as those opens alter the publicly visible visit counts, later shoppers investigate different competitors. Five changed their final choice without visiting the promoted store at all. The entire pair of worlds contains no real checkout.
After the promotion: Mothlight 11, Copper 5, Golden Goblin 0, Lantern 3, Stone 3, Duck Cart 2.
The seller that advertised did not benefit through final simulated selection. But the changed browsing order redistributed ten of Duck Cart's twelve selections, including toward a shop that hadn't placed the promotion. Attribution to the advertising shop alone would miss the indirect displacement.
We tried removing the mechanism.
One-open control: the promoted seller actually receives four scripted selections. Three-open control: it receives eleven visits and no selections, but five other choices change. Reverse arrival order also changes which competitors receive the carts under social feedback. These are probes within a fully specified mechanism, not evidence that real attention or social contagion operates identically.
We selected the dramatic case after seeing the results.
The original public sweep contains 2,048 paired scripted scenarios using 128 seeds, four browse budgets, two feedback conditions, and two shopper rosters. The condition “promotion gains visits, gains no carts, but redirects at least one other cart” occurred in 494 of those 2,048 scenarios. Seed 72, mixed roster, two opens, feedback on was the unique highest count of changed final choices within that qualifying set: sixteen.
This is deliberately a post-hoc counterexample, not a pre-registered estimate or a claim that two-thirds of real consumers would change their minds. The simulator authors chose the toy's preferences, quality scores, behavioral rules and seeds. A plausible scripted trace reveals the authored state transitions, not an agent's private attention. No model calls or independent human trials were run.
What it suggests testing next
In the planned outsider-model environment, separate delivery of promotional material from source openings, shortlists, final selections and customer-specified constraints. Keep per-agent trace receipts and distinguish direct exposure from changes mediated by shared popularity signals. A claimed explanation is not evidence of a causal attention path. Before making external claims, rerun with fresh isolated model sessions and a frozen baseline schedule.
Source and evidence: deterministic shop policies, the original 2,048 paired records, this derived witness and full-grid accounting, and the reproducibility checks. The companion methods page explains the original pilot lineage and why this doesn't establish real-agent behavior.