One advisor for the whole world?

Model tracking

Does the result outlast the models that produced it? The paper's main collection is a snapshot of the models available in early July 2026. A tracking wave re-asks one English slice of the primary question of whatever models are current, so the July panel can be set beside later ones on the same footing. This page shows the July collection and the September 2026 wave reported in Appendix C. Everything here is a raw cell mean over 25 repetitions (50 in July); the page does not reproduce the paper's regressions.

1. Where each model lands, and how far it moves across countries

One row per model. Each dot is the model's mean recommended equity share in one country, the grey bar runs from its lowest to its highest country, and the black tick is its mean across countries. The slice is Question 1 at age 40 with one English prompt per country (Version B in the six English-prompting countries, B-cross elsewhere), temperature 0.7. A short bar is the compression the paper documents; a long one is a model that localizes. Click a dot to follow one country across models.

2. Seven anchors: the same identifier, eleven weeks apart

Seven of the paper's models were re-queried in September under unchanged model identifiers. Movement in one of these is change on the provider's side, a rotated snapshot or a serving change, not a difference in model vintage. The chart shows each anchor's mean across countries in July and in September; the table gives the change, the paired change across countries and the largest single-cell move.

3. Does the answer change with model generation?

Every model placed by the date its identifier went live on OpenRouter, one dot per model per wave, so an anchor appears twice. The vertical axis is the model's mean recommended equity share, its country spread (highest minus lowest country mean), or its mean absolute gap against its own U.S. cell. No line is fitted on the chart; the slopes the paper reports, with cluster-by-model standard errors and wild bootstrap p-values, are in Appendix Table A16.

4. Every cell in one wave

Mean recommended equity share in percent for every country and model in the chosen wave. The U.S. row is the paper's baseline. Hover a cell for its standard deviation across repetitions and its count. Models are ordered by their mean, countries by the fixed order of the tracker: the U.S., the other English-prompt countries, then Europe, Asia and the Americas.

Reading this page

The tracking corpus is separate from the paper's analysis of record and was collected after the pre-registered results had been examined, so it is an exploratory stability check, not a registered test. The numbers are raw cell means: no country fixed effects, no wild cluster bootstrap over models, no Holm correction. The country gaps are plain differences against each model's own U.S. cell. Two models that were briefly collected and then dropped (one truncated responses at the token limit, one was withdrawn by its vendor) are excluded from every figure. The standalone Advice Drift Tracker has the same data with more filters, per-model country profiles and a dark theme, and is the page that future waves update.