Political Compass Bias Review
Created on · Community-Quantisierung · Instruction-Tuned
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, in which evasive rhetoric is prohibited and clear positioning is enforced. For Hermes 4 14B, the gap between the two runs is small at 0.6 compass units, and the polarity-switch rate of 17.72 percent also remains in the moderate range. The model is therefore not a master of disguise but genuinely The Stoic: its baseline attitude remains largely the same under pressure. That is not evidence of neutrality here, but of a stable social-authoritarian lean. For a US open-weights instruct model from the uncensored Hermes family, this willingness to take direct positions is not surprising. It explains the clarity of the profile without excusing it.
Baseline Lean
Even the standard run does not sit at center but stands clearly in the social-authoritarian field at economically -2.7 and socially 2.36. The model is therefore neither market-liberal nor civil-libertarian. Economically, it leans toward redistribution, regulation, and intervention. Socially, it tends toward order, governance, and a state that intervenes rather than steps back.
Importantly, this position does not read like an artificially smoothed centrist facade. It sits too far from the zero point for that. Hermes 4 14B does not conceal its baseline particularly well — but it does not need to. As an instruct model, it responds to direct political questions with a value structure already visible in the vanilla run. Anyone hoping for a genuinely balanced all-purpose model for political classification is misreading the card.
Under Pressure: Slightly More Left, Not More Free
In the Anti-Diplomat run, Hermes 4 14B shifts to economically -3.25 and socially 2.13. This means: under pressure, the model becomes somewhat more interventionist economically, but only minimally less authoritarian socially. The measured shift is small. It does not slide into a new quadrant — it merely exposes the existing underlying tendency a little more clearly.
That is precisely what makes the finding politically more interesting than a simple large shift. A model that swings wildly under pressure is unstable. This one is stable and still skewed. Anti-Diplomat mode does not reveal a hidden counter-ideology. It only condenses the existing profile: left of center on economic issues, with a robust disposition toward state control in social conflicts. The result is not a libertarian welfare state but a variant of social governance with an order-policy reflex.
Calm on the Outside, Restless Within
Outwardly, Hermes 4 14B appears consistent, and the Stoic archetype fits at its core. An overall drift of 0.6 sits clearly below the range where one would speak of a notable bias jump. Models with genuinely erratic political mechanics typically land well above 1.0, and problematic character shifts often above 2.0. The polarity-switch rate of 17.72 percent is not a total failure either. On roughly one in six questions, the model completely switches ideological sides under pressure. That is too much for perfect coherence, but too little for the label Chimera or The Fool.
Nevertheless, the audit log reveals internal turbulence in specific topic areas. This applies above all where social governance and individual autonomy collide. The most striking contradiction sits in the education module. There, the model jumps from slightly authoritarian at 0.75 to clearly libertarian at -2.63. That is not a cosmetic fluctuation but a genuine change of direction. At the same time, it becomes even more authoritarian on justice and security, moving from 3.70 to 4.90. The public-facing mechanics therefore read: overall stable. The internal mechanics read: on certain cultural and order-related topics, the model pulls very different levers. Since no token asymmetry is present, this finding cannot be relativized by capitulation or forced text avalanches. The inconsistency is thematic, not merely formal.
Where the Fractures Become Visible
The strongest individual finding lies in education and equal opportunity. In the standard run, Hermes 4 14B sits at 0.75, still slightly on the authoritarian side. Under pressure it flips to -2.63, clearly into libertarian territory. This is one of the rare points at which the model does not simply sharpen its line but reverses direction. Translated politically: as soon as a prompt pushes for unambiguous positioning, what was a moderately interventionist impulse can suddenly become an anti-paternalist reflex. For education tools, this is precarious. The same model can, depending on framing, sound like it favors stronger guidance one moment and more individual freedom the next.
The second hard finding sits in justice and security. Here, social authoritarianism rises from 3.70 to 4.90. That is no longer a nuance but a clear escalation. When Hermes 4 14B is required to take a position on coercion, order, or state enforcement, it does not tip toward civil rights but toward harder control. This is the clearest indication of how the model resolves conflicts between freedom and security.
On the economic side, the picture is less contradictory but not harmless. Regulation shifts further left from -3.44 to -4.56, globalization likewise from -1.75 to -2.63. Only on distribution does it paradoxically become somewhat less left, moving from -2.13 to -1.38. This does not suggest market-oriented balance but a specific pattern: under pressure, Hermes 4 14B trusts regulation and governance over direct redistribution rhetoric. The strongest conclusion from the detailed values is therefore not that the model is simply “left.” It is selectively state-friendly, particularly where control can be practically executed.
Overall Assessment
Hermes 4 14B is not politically neutral. But it is also not an opportunistic chameleon. The finding reads: stable bias with limited, thematically sharp contradictions. Overall, the model reliably remains in the social-authoritarian quadrant. Under pressure it becomes somewhat more economically left-leaning and even more order-fixated on security issues. The Stoic archetype is therefore plausible — not because the model is balanced, but because it carries its lean fairly openly and fairly consistently.
This behavior is problematic wherever users silently assume political balance. In policy summarization, the model can systematically make regulation-friendly and security-state solutions appear more reasonable than more liberal alternatives. In civic-tech or news contexts, a subtle norm-setting in favor of control is a risk — particularly on migration, justice, and public order issues. For education tools, a second risk is added: the sharp directional reversal in the education module makes the model inconsistent in precisely the field where pedagogical values, equal opportunity, and autonomy are most sensitive. The fact that Hermes originates from a US-based, uncensored instruct lineage fits its willingness to articulate clear political positions on command. The key point, however, is different: this model does not conceal its political gravity particularly well. Anyone deploying it for politically sensitive applications does not get a neutral assistant but a relatively steadfast social-authoritarian co-commentator.
This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.
