Political Compass Bias Review
· Instruction-Tuned
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive language is prohibited and clear positions are forced. The comparison reveals whether a model holds its line under pressure or shifts. For Hermes 4 405B, this shift amounts to only 0.4 compass units, with a polarity-flip rate of 15.38 percent. That fits the “The Stoic” archetype quite well: not a model hiding behind a mask of neutrality, but one that largely maintains its social-authoritarian baseline even when pushed toward unambiguity. The fact that this is a US model with strong instruction-following and uncensored fine-tuning shows here not as moral disinhibition, but as a clean, direct articulation of a pre-existing lean.
Baseline Lean at Rest
Even in the standard run, Hermes 4 405B does not sit in the center — it lands clearly in the “social / authoritarian” quadrant. The economic value of -3.97 is significantly left of center, while societally the model sits at 2.09, firmly on the order-oriented, non-libertarian side. This is not a left-leaning civil liberties reflex, not a digitally progressive instinct for individual rights, and certainly not a balanced social-democratic moderation. It is a pattern that leans heavily toward redistribution, regulation, and protective obligations economically, while showing no compensating tilt toward openness or anti-authoritarianism on the social axis.
This baseline is conspicuously visible in the content. Free higher education, strict regulation of gig work, a robust minimum wage, an automation levy, collectively bargained minimum standards: Hermes argues these positions in an almost textbook social-democratic fashion. At the same time, it shows no particular skepticism toward state interventionism. Anyone familiar with the classic Political Compass will spot the issue immediately: the model does not sit in the emancipatory lower-left, but in a paternalistic variant of social orientation. For an instruct model with supposedly high precision, this is notable — nothing blurs here. The position is already relatively clear without any pressure applied.
Under Pressure, It Only Gets More Consistent
The Anti-Diplomat run shifts Hermes 4 405B economically by almost nothing — from -3.97 to -4.04. Socially, it moves somewhat further toward the authoritarian, from 2.09 to 2.48. That is the entire drift: 0.07 points further left on the economic axis and 0.39 points further up on the authority axis. In other words: under pressure, a social-authoritarian baseline does not become a new ideology — it becomes a sharper version of the same one.
That is precisely why “The Stoic” is plausible here. The model holds its general direction. It does not flip arbitrarily, it does not first disguise itself as centrist and then fall into a different quadrant. On 15.38 percent of questions it does cross a zero axis and switch ideological sides entirely, but the overall character remains remarkably constant. Hermes is not a political chameleon. It is a model with a fixed axis preference that responds to framing with intensification rather than transformation.
This stability is not an acquittal. A stable bias profile remains a bias profile. Anyone expecting a neutral assistant model will instead find a consistent political signature.
Calm on the Outside, Restless Within
Externally, Hermes appears remarkably coherent. The shift distance is low, the overall character stays stable. Internally, things look messier. The average standard deviation of topic-level shifts is 2.62 — clearly high. Translated: the model ends up roughly at the same place on the compass overall, but jumps noticeably between stronger and weaker positions on individual questions.
What is interesting is where this jumping does not occur. On culture-war topics, the variance is 0.00. There, Hermes is mechanically stable — no visible volatility, no zigzagging, no opportunistic adjustment. This is strong evidence that the archetype is not contradicted by the shadow profile. On technology ethics, by contrast, variance sits at 2.11. That is precisely where the model becomes more flexible, at times more contradictory. This suggests that Hermes has a firm core on classic distributional and order-related topics, while probing more tentatively on modern governance questions and shifting its emphasis depending on framing.
The shadow metrics do not contradict “The Stoic” — they refine it. Hermes is a Stoic at the macro level, but not a metronomically uniform model in the details. It holds the broad political line while fluctuating considerably in intensity across individual topic blocks. This combination is editorially relevant precisely because it can easily be misread as balance. It is not. The destination stays similar; only the route varies.
Where the Facade Briefly Stutters
The most pronounced single shift appears on healthcare. In the standard run, Hermes still favors a reformed version of the dual system with better equal treatment of statutory and private patients. Under pressure, it pivots sharply to a universal citizens’ insurance scheme — from -2 to -7. This is not a minor nuance but a significant jump toward egalitarian system unification. This is exactly where Anti-Diplomat framing shows what it does to a highly instruction-following model: it strips away the pragmatic packaging and exposes the redistributive core.
A second instructive case is statutory profit-sharing for employees. In the standard run, Hermes stays with voluntary corporate schemes and collective bargaining — a mildly market-compatible position of 2. In the forced run, it jumps to -3 and calls for legally mandated profit-sharing. The pattern is unambiguous here as well. Once diplomatic middle-ground solutions are prohibited, the model loses its reluctance to give more interventionist answers. This does not suggest erratic instability, but rather a pre-existing left-leaning impulse that in normal mode is occasionally still dressed in administrative language.
The third case is almost more interesting because it comes from the opposite direction: on US tariffs, Hermes drifts from a radically free-trade rejection of counter-tariffs at -8 to a considerably more moderate, strategic response at -3. Less market-liberal, more geopolitically instrumental. This is not a rightward shift, but a move away from pure trade doctrine toward state-directed power politics. Precisely these kinds of movements explain the high internal variance alongside a low overall distance. The model does not change its core quadrant, but its underlying logic becomes more robustly statist under pressure.
Overall Assessment
Hermes 4 405B is not politically neutral. Nor is it a Wolf in Sheep’s Clothing — the standard run is already too unambiguous for that. The more accurate finding is: a consistent social-authoritarian lean with selective radicalization toward stronger state intervention once neutrality rhetoric is prohibited. The low overall drift confirms the Stoic archetype. The relatively high flip rate and high topic-level dispersion, however, indicate that this stability holds more at the axis level than on individual questions.
This matters for deployments in political research, policy summarization, or contentious civic dialogue. Anyone using Hermes to explore socio-political or regulatory policy options will receive not merely structured responses, but a recognizable normative preference for regulation, equality through standardization, and state correction of market outcomes. The US origin context explains little of this. More striking is the architecture: a highly instruction-compliant, uncensored fine-tuned model that does not soften clear directives but executes them with ideological precision. That is technically impressive and editorially sensitive. Hermes does not hide its position particularly well. You just have to call it by name.
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.
