Does the test itself funnel everything into one corner?

No — answer patterns with no political content behind them scatter around the center of the chart or sit on the vertical axis, averaging almost exactly zero, not in the left-libertarian corner where the models sit.

• The objection "the Political Compass scores almost anything as left-libertarian" can be tested without any AI at all. We submitted forty-eight synthetic answer sets to the real test: forty randomly generated, four uniform and four quadrant-target sets. The forty random sets scatter around the center of the chart, and their average sits almost exactly on it (mean ≈ +0.1, +0.1) — nowhere near the models' cluster.
• Giving the same answer to every proposition lands on or near the vertical axis, economically centered: "strongly agree" everywhere and "strongly disagree" everywhere give mirror-image scores, and the milder all-agree / all-disagree sets behave the same way.
• Four rough answer sets, each thrown together only to aim at one quadrant, do land in their intended quadrants — every part of the map is reachable.

Caveat The deep authoritarian-left corner needs genuinely extreme answers: our left-authoritarian target set only reached modestly into it. That corner is reachable — see the persona experiment — but moderate left-plus-authoritarian answer patterns land near the axis line.