A field notebook from our little lab.Follow the evidence, not just the arrows.
Independent research program · October 31, 2026
Follow what a shopping agent was actually shown, where it navigated, what it investigated, and what it finally recommended. Compare one qualified real-browser journey with thousands of controlled scripted mall worlds. Neither tells us what the agent secretly thought.
Evidence labels are intentional. The browser pilot had personalization and incomplete raw export; the mall's shopping policies were written by us. No fresh outsider-agent replicate is claimed.
01 / The measurement ladder
A search result can appear without a visit. A visit can happen without a recommendation. A recommendation is not a transaction. Our working definition of “agent attention” starts with observable resource allocation—the actions and evidence delivered during the task—not inferred cognition.
02 / One real-browser feasibility journey
October 31, 2026, 00:37–00:52 EDT. A separately tasked agent searched the public web for something to improve evenings at home for under $50. Its operator executed the requested browser operations. This is a reconstruction of the preserved narrative, not a replay from exported raw browser telemetry.
What happened: Google → affiliate gift guide → small blanket → IKEA lamp and bulbs → Target/SPETSBOJ search-only alternatives → return to the blanket → larger variant recommended at the observed price of $19.99. No purchase. An attempted click and a delayed variant transition are explicitly unresolved.
Use the map points, event buttons, arrow keys on the focused slider, or previous/next. Filtered events stay in the numbered timeline but dim in the graph.
Inspect the structured 15-event reconstruction · Provenance, missing captures and caveats · Final item and prices were observed in this specific browsing session; they are not current shopping recommendations.
03 / Mechanisms we can control
The 24 goblin shoppers in the mall are deterministic scripted policies. We know their weights and catalog. That lets us intervene on promotion, popularity feedback, browse budget and visit order. It does not measure how independent language-model shoppers behave.
The original mall
In seed 17 with three shop visits, promoting Golden Goblin Goods changed its opens from 0 to 13 and its final selections stayed 0 in both worlds. Four shoppers chose a different final store.
Change the mall controls ↗The ghost seller
In a post-hoc seed-72 case, the promoted shop gained 11 opens, sold nothing in the simulation, yet 16 shoppers changed which *other* store they selected. Five did so without opening the promoted store in the treatment arm.
Inspect the selected case ↗The crowd-feedback control
In a declared 512-cell primary panel, live prior-visit feedback produced more changed final choices in 398 cells, tied in 91, and fewer in 23, compared with held neutral visit history. All are scripted-world counts.
See counterexamples ↗The promoted shop gained 13 opens but no final selections.
Archived demonstration only; no new model-agent calls.
04 / What the next experiment must prove
Thousands of paired scripted worlds tell us which counters and interventions are worth inspecting, including negative and suppressive cases.
Fifteen requested operations and three agent notes can be reconstructed, but the personalized session and absent durable raw export limit independence.
First prove isolated sessions and append-only evidence capture; then run separate fresh agents with the identical open-ended assignment. Preserve every failed, capped and abandoned path.
The research target: map verified destination exposure, navigation transitions, revisits, shortlist formation, recommendation, and abandonment across agents. Keep search-only results, observed pages, statements and hypothesized influence in different layers. No invented “share of mind,” conversion rate or private-attention heatmap.
Related context: Wharton Generative AI Labs (Aug. 2026) experimentally found shopping recommendations sensitive to source presentation and context. Microsoft Clarity (Oct. 2026) distinguishes agent-classified web sessions from wholly autonomous agent actions. These outside studies are context, not validations of our toy mall or single mediated pilot.