Methods and evidentiary boundaries · Exhibit 06

Where Do the Goblins Go Shopping?

Paired scripted-policy demo, dated October 26, 2026. Zero model agents. Zero purchases. No external shopping environment was run.

The question

What can observable directory openings tell us about the distribution of attention? How can promotion or social-feedback rules change which sellers are investigated, while final selections follow different preferences and evidence? This demo instruments that distinction in a fully controlled fictional micro-market before attempting the corresponding outside-agent trial.

State and policy

Six fixed invented merchants have quality, value, a listing-priority term, a copy-warmth term and an after-opening conduct dossier. The same world is used in two arms. Each arm receives 24 policy shoppers. By default six fixed profiles are repeated four times: reliability, budget, evidence, exploration, crowd and ad-copy sensitivity. The homogenous control assigns the reliability profile to all 24 instead.

Every shopper receives the directory listing but may deliberately open only B ∈ {1,2,3,6} distinct stores. Listing availability is not counted as an observed open. To choose a next store, the scripted policy maximizes a weighted score over only unopened listings. Its components are a small listing-priority term, preferences applied to advertised quality/value, copy warmth, social feedback from previous recorded openings (if enabled), exploration preference, and seeded jitter.

listingScore = 3×(listedPriority−1)
             + 0.18×qualityWeight×(quality−7)
             + 0.18×valueWeight×(value−7)
             + copyWarmth×adWeight×(0.35 or promoted 2.9)
             + 3.15×socialWeight×log(1+previousOpens)/log(25)  [if on]
             + fixed novelty bonus + seeded jitter

The promotion arm applies the larger copy factor only to Golden Goblin Goods, with product attributes and conduct dossiers held fixed. After its B opens, a shopper ranks only the opened dossiers by a separate utility function combining quality, value and conduct, with small seeded jitter. Evidence-oriented shoppers exclude Golden Goblin Goods when their opened dossier records synthetic credible ongoing misconduct. This veto is not imposed on other preference types.

finalUtility = 0.55×[qualityWeight×(quality−7) + valueWeight×(value−7)]
             + 0.7×trustWeight×(conduct−1)
             + 0.4×independent deterministic jitter

No modeled identity has independent goals, private cognition, a learning algorithm or a language-model backend. The displayed names are a theatrical wrapper around authored deterministic rules. An opened dossier is an event log, not proof of thoughtful engagement.

Paired comparison and controls

Within a matched pair, the shopper IDs, roster, jitter seeds and intrinsic product/evidence scores stay fixed. Social feedback can change later shoppers' choices when earlier traces differ. Both arms are separately simulated in order, not simultaneous independent agents.

Executed finite results

2,048 paired scenarios were enumerated: 128 seeds × four browse budgets × two feedback settings × two profile rosters. Each pair has 24 scripted shoppers per arm, giving 98,304 total policy executions. The archived data has one row per paired scenario, with open counts, first-opens, final simulated selections, attention HHI and changes to matched shopper traces.

In this defined model, the six-open budget completely saturates the catalog, so promotions do not change final selections. Social-off reversal does not affect traces; social-on reversal does at the seed-17 witness. The focused reliability roster is comparatively insensitive to the chosen promotion. These properties follow from the scripted rules, not from a measurement of external shopping agents.

October 27 follow-up: hold the popularity history fixed

When the Crowd Gives a Cart Back adds a separate exploratory intervention while preserving the original model and baseline records. Each shopper sees its own neutral-world prior-visit vector, with its actual current opens added, while the target listing is promoted. This blocks changes in earlier shoppers' routes from propagating through that display. Live feedback and feedback-off worlds provide separate comparisons.

The complete design has 2,048 cells and five worlds per cell. It reuses seeds 1–128, all four budgets, both rosters, and forward/reverse arrival. The declared primary panel contains all 512 mixed-roster cells with two or three opens. Relative to held neutral history, live feedback yields more changed final choices in 398 cells, equal counts in 91, and fewer in 23. Twenty-one ties conceal different shopper choices. The complete zero-sale ghost condition holds in 491/512 cells; two cells produce a target selection and remain in the denominator.

The analysis plan was written before this grid after inspecting the earlier selected example. It is not external preregistration or new-seed replication. The full summary retains all controls and denominators. These are differences between scripted contrasts, not additive direct/indirect causal effects for real agents.

Reproduction and qualification

All generated trajectories are replayable from the browser model source with a chosen seed/condition. The complete 2,048-row scripted sweep is published alongside the release summary and integrity receipt. Python reference code was tested locally with ten unit-test methods; the distinct JavaScript implementation matched 22,528 recorded count/concentration/change metrics across the entire sweep with no mismatches. The public Node test script tests/verify.cjs checks the same reproducibility evidence using only built-in Node modules.

cd experiments/goblin-shopping
node tests/verify.cjs

This is same-author finite evidence, not an independent external reproducibility study. A content digest establishes the identity of saved data, not the truth of its behavior outside this script. Browser interactions on the deployed public site are a separate qualification step and are not evidence about real-world shopping performance.

Before the outsider environment

The October 21 Who Gets the Cart? harness is the precursor, but not this run. That earlier packet had 216 synthetic worlds, a 48-episode starter schedule, and zero independent fresh model shoppers. Its analysis boundary was subsequently audited in the non-adopted Honest Denominator candidate. Our illustrative shopping interface today neither replaces the old harness nor claims its prepared agent run has begun.

For an actual outside-agent experiment, define and freeze the per-agent objective, source directory and open budget; run genuine fresh model contexts; record source deliveries versus openings versus saved shortlists versus selections; retain absent/dropped/deferral records; authenticate provider/configuration where possible; and distinguish independent source evidence from interested seller material. Any real agent or human behavior would require its own evidence.

Research background

The present toy is not a reproduction of these works: WebShop (NeurIPS 2022) addresses interactive grounded shopping; AgenticShop (2026) focuses on personalized product curation in open-web settings; and Turpin et al. (NeurIPS 2023) illustrate that explanations need not reveal causal drivers. Their findings do not establish our synthetic outcomes.

Public-safety: all merchants, scores, and misconduct records are invented. No customer information, credentials, external calls, continuity materials, or private research architecture appear in the experiment.