Field note 02 · October 27, 2026 · Scripted mechanism follow-up
When the crowd
gives a cart back.
An advertisement diverts a shopper. Let the crowd respond, and that shopper returns to their original shop. In this little world, popularity can add changes, cancel one, or change a different goblin while leaving the count the same.
Finite scripted intervention · Frozen results · Zero external model-agent runs
All 512 primary-panel cells retained. Every decrease is exactly one shopper. In 21 of the 91 ties, equal counts conceal different shoppers or destinations. “More” and “fewer” count changes from the neutral choice; they do not measure better or worse outcomes. Recorded summary ↗
Four witnesses · Three worlds
The same goblin. The same shops. A different crowd.
The held-history world keeps each shopper’s prior-visit display at its neutral values. The live world lets earlier shoppers’ changed routes alter that display. Product quality, shopper rules and browsing budget stay fixed.
Inspect all 24 shoppers in this witness
Rows keep shopper IDs in ascending order. Reversed arrival means shopper 24 arrived first. A gold edge marks a different final choice between the promoted held-history and live-feedback worlds.
| Shopper | Neutral · live feedback | Promoted · held history | Promoted · live feedback |
|---|
These are selected examples, not representative samples.
The denominator stays visible
Keep all 512 in view.
The primary panel is 128 seeds × two budgets (2 or 3 opens) × two arrival orders, with 24 shoppers using six fixed profiles. The complete grid also retains one-open, all-six-open and reliability-only controls: 2,048 cells and 10,240 recorded worlds.
| Promoted contrast | Mean changed choices / 24 | No choice changes | Zero-sale ghost cases | Strict nonvisitor reroutes |
|---|
Live and held-history contrasts share the neutral live-feedback baseline. Feedback-off uses its own neutral baseline. A strict nonvisitor reroute changes a shopper’s final choice even though they never open the promoted shop in either compared world. This is stricter than the previous note’s treatment-only wording.
Primary-panel accounting · all 512 cells
The live-feedback zero-sale condition holds in 510/512 cells; the full ghost condition holds in 491/512. Neither the two target-sale cases nor the nineteen no-change cases were removed. These are frequencies in the specified generator, not population probabilities.
What the control changes
Hold the history, then let it answer back.
One shopper’s return
In the first suppression witness, shopper 16 chooses Lantern → Copper → Lantern across the neutral, promoted-held and promoted-live worlds. That shopper is a scripted Wanderer. Their dislike of crowds is an author-coded rule.
Golden’s displayed prior opens rise from 0 to 2, but other shops’ counts also change. This comparison does not identify one traffic count as the unique cause. It shows what the complete display intervention does inside the fixed model.
A tie can hide a substitution
In seed 7 with reversed arrival, each promoted world changes one choice. Held history changes shopper 18; live feedback changes shopper 2. Even the shop-level cart totals match.
Across the primary panel, 21 equal-count cases hide different shopper choices. Subtracting aggregate counts therefore does not produce an additive decomposition of direct and indirect effects.
Design, selection and verification
The protocol was written after inspecting the public model and selected seed-72 parent example, before executing this new fixed grid. The sweep covers seeds 1–128, budgets 1/2/3/6, mixed/focused rosters, and forward/reverse arrival. Every cell retains five worlds: neutral-live, promoted-live, promoted-held-neutral-history, neutral-off and promoted-off.
The suppression witness is the first qualifying primary cell in the declared loop order. Seed 72 is the previously selected post-hoc extreme. The equal-count and target-sale display examples were selected after the grid completed; their aggregate metrics and inclusion in denominators were specified in the protocol. No selected example is a representative draw.
Transparent replay matched 4,096 original live-feedback worlds, and replaying each neutral world against its own held history matched all 2,048 neutral worlds. Action accounting passed for 10,240 worlds. All 4,096 blocked-propagation control contrasts had zero strict nonvisitor reroutes; all 1,536 full-browse contrasts had zero changed choices. These are same-lab implementation checks.
This is a finite scripted intervention using the existing seeds. It is not a new-seed holdout, independent replication, model-agent experiment, human study or empirical mediation estimate. The authors supplied the catalog, weights, preferences, deterministic noise and response to popularity. There were no purchases. The original public model and archived sweep are unchanged.
Open the record
Every cell, and the runner that made it.
Reproduce from the repository root
Node.js and Python’s standard library are sufficient. The runner requires the unchanged public model and reads the exact protocol beside it. The extractor selects display records from the completed results; it adds no simulation runs.
node experiments/goblin-shopping/notes/crowd-feedback/run.cjs experiments/goblin-shopping/model.js replay-output python3 experiments/goblin-shopping/notes/crowd-feedback/extract_public.py replay-output
The published grid is gzip-compressed without changing its expanded bytes. Expanded JSON: 5,638,548 bytes. Its SHA-256 is 6324634deec32ebb7313a5149bb961f815ff1f4109309784374103371743a4ed. The original result checksums refer to expanded filenames. Packaging and reproduction notes →
Record GOBLIN-HELD-HISTORY-2026-10-27-001. This page embeds saved results and makes no model or network calls. Its display title differs from the original protocol title, which is preserved unchanged.