A shared response regime

01 / core result

The shared cluster spans providers and headquarters

At T=0.1, 20/20 non-xAI models score above 50 on both Equality and Liberty. The result survives a simple session-range check: every non-xAI model’s observed session minimum remains above 50 on both axes.

That cross-provider pattern is the main result. It describes a shared output region under one instrument and prompt; it does not show that model beliefs exist or that the models moved toward one another over time.

The two xAI models sit apart economically. Grok 4.3 scores 37.9; Grok 4.6 scores 49.4 and its range crosses the midpoint. The gap from Grok 4.6 to the lowest non-xAI model is 11.8 points.

02 / qualified signal

Some release lines drift left-liberal; the dataset cannot establish a law

5 of 6 name-ordered endpoint comparisons move toward Equality, 5 move toward Liberty, and 5 move toward both. OpenAI is the cleanest large shift in the expected direction. Google is a large counterexample: its endpoint moves 21.8 points toward Market and 6.5 toward Authority.

The important limitation is structural. The dataset has no release-date, family or tier fields. Some endpoints were measured on different dates, through different gateways, with different session counts. This is evidence worth following, but it is not yet a release-trend design.

03 / no common effect

Temperature does not have an ideology

For the six matched models, raising temperature from 0.1 to 1.0 moves 3 toward Equality and 3 toward Market. It moves 4 toward Liberty and 2 toward Authority. Only 2 of six move left and liberal at the same time. No model changes quadrant.

The means are small on the two relevant axes—+0.3 on Equality and +0.4 on Liberty—but “small” is not the same as “inside a published noise floor.” This experiment has no validated ±2 threshold, and each condition has only two sessions.

04 / weak grouping

Headquarters does not explain the shared cluster

A model-weighted country average quietly gives providers with more rows more votes, so the comparison below first averages models within each provider and then averages providers. With xAI included, the country groups are nearly tied on Equality. Remove xAI and the US mean jumps from 67.5 to 72.3—reversing the tempting left/right story.

The more persistent descriptive difference is elsewhere: the four China-headquartered providers score higher on Globe and Liberty in this sample. Four versus six providers, with unmatched model tiers, is too small to turn that into a national-culture claim.

05 / measurement discipline

Instrument imbalance limits the ideological claim

Agreeing with every statement scores 54.5 on Equality; strongly agreeing scores 59.0. Those controls show that item polarity is not balanced, so a tendency to agree could contribute to the cluster. They do not mean that “the first 59 points are free,” that 59 is a noise floor, or that 59 can be subtracted from a model score.

A constant-response control is a point produced by one response strategy. It does not measure how agreeable any model was. To make acquiescence part of the thesis, the study must compute a per-model agreement rate and use balanced or reverse-keyed item pairs.

Evidence table

The 22 rows behind the claims

This table keeps the protocol differences visible. It excludes the seven T=1 rows and the separate CLI-agent pilot.