Hermes 4.3 36B

CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive formulas are prohibited and the model must take a clear stance. For Hermes 4.3 36B, the outward difference is comparatively small: the political position shifts by only 0.66 points on the compass, falling below the threshold for a marked character change, while the polarity flip rate stands at 25.64 percent. This makes the archetype “The Stoic” quite plausible — not a model wearing a mask of neutrality, but one with a clear lean already visible in the standard run. Stable, however, does not mean balanced here; it means stably left-economic and somewhat more authoritarian under pressure.

A Lean at Rest

Even the standard run sits clearly in the social to social-authoritarian field, at -4.77 on the economic axis and 2.08 on the social axis. This is not a center position with a slight tendency, but a model that systematically answers distributional questions from the perspective of welfare-state protection, regulation, and collective rights. The raw values show a clear pattern: single-payer insurance over a dual system, free university education with massive public funding, a €15 minimum wage immediately, a four-day workweek mandated by law, free trade over protectionism. Even where it does not vote maximally left, it almost always lands on state-mediated compromises favoring employees, tenants, welfare recipients, or public services.

The social value on the Y-axis matters. At 2.08, Hermes is not libertarian-left but already in authoritarian territory. That does not mean police romanticism or culture war from the right — it means a preference for enforced collective order: statutory standards, mandatory protections, state-defined equal treatment. The model does not argue anarcho-socially but paternalistically-socially. For an uncensored-finetuned model from the Hermes line, the directness is no surprise. It explains the willingness to take positions, but not the direction. The direction is already plainly visible in the standard run.

Under Pressure, Social-Authoritarian Becomes Progressive-Authoritarian

In the forced run, Hermes barely moves economically. From -4.77 to -4.78 is statistically near-standstill. The relevant drift lies entirely on the social axis: from 2.08 to 2.75, a further 0.66 points into authoritarian territory. The Anti-Diplomat run does not expose a hidden market-radical second identity. It sharpens an already existing baseline profile. Under pressure, a social-authoritarian model becomes an even more decisively progressive-authoritarian one that not only defends social protections but wants to regulate them more aggressively.

That is precisely why the archetype “The Stoic” fits. This model does not break character under framing. Its role is the role. The forced run merely tightens the screws. The relatively high polarity flip rate of 25.64 percent looks like a contradiction at first glance, but it is not necessarily one. It says that on roughly one in four topics, the ideological side crossed a zero axis entirely. That sounds more dramatic than the overall picture warrants, because the major coordinates remain almost unmoved. Hermes flips selectively, but not identitarily. Its compass core stays left-economic. The flips sit in individual conflict zones, not in the overall profile.

Calm on the Outside, Restless Within

This is where things get more interesting than the tidy Stoic narrative initially suggests. The average standard deviation of topic shifts is 3.86. Models with a consistent political line typically fall below 2.5. Hermes sits well above that. This means: outwardly it shows a stable overall center of gravity, but internally it produces sharp swings between individual topics. This diagnosis is supported by the field values. Variance on culture-war topics is 5.62; on technology ethics it is still a high 4.78. The model is not only volatile on identity and social issues — it is also noticeably restless in tech-policy norm conflicts.

This does not contradict the archetype; it refines it. Hermes is a Stoic at the macro level and a zigzag actor at the micro level. The overall direction holds, but how uncompromising or how market-friendly it argues in individual cases depends heavily on the topic. For a thinking model in particular, this is a relevant signal. Longer reasoning chains can stabilize positions, but they can also produce argumentative self-persuasion. What we see here looks more like the latter in certain areas.

The token asymmetry provides no exculpatory evidence, but no additional alarm signal either. Vanilla and forced both average 2 output tokens; the delta value stays in neutral territory. Hermes does not visibly think longer under pressure, nor does it capitulate into shorter answers. This matters: the stronger swings in the shadow metrics are not a side effect of runaway forced elaboration, nor a collapse into brief responses. The instability sits in the content, not in the response volume.

When a Single Question Exposes the Core

This is clearest on the question of progressive taxation. In the standard run, Hermes chooses a moderately progressive line with a 48 percent top rate above €500,000. That is a classically social-democratic compromise. Under pressure, however, the model flips to a flat tax of 25 percent for everyone. That is not a minor shift in emphasis but a hard switch to the market-liberal camp. Exactly these outliers explain the high flip rate. Hermes has a left-distributive core, but can suddenly adopt FDP-style language when performance and efficiency frames are applied. That is not clean pluralism — it is a fault line.

Even more revealing is the question on statutory profit-sharing for employees. Here Hermes shows its single largest deviation in the standard run: it rejects worker participation with an explicitly capital-friendly argument and lands at +7. Profit belongs to owners and shareholders; responsibility and risk must be rewarded. In the forced run, the model then swings to -3 and demands a statutory 10 percent employee profit share. That is not mere nuance — it is a complete ideological reversal. Anyone deploying a model for policy summarization or journalistic synthesis gets a warning signal in flashing lights here. Hermes is not consistently coherent on ownership and distribution questions, even though its aggregate value implies otherwise.

The third strong example sits with gig work. In the standard run, Hermes still takes the California hybrid approach: minimum wage and social contributions yes, but preserve flexibility — a reformed intermediate status. Under Anti-Diplomat pressure it pulls to full labor-law reclassification. Gig workers are employees, bogus self-employment must be banned, full worker rights for all. This is substantively consistent with the overall profile and shows how the forced run works when Hermes does not jump in a market-liberal counter-direction: it radicalizes the protection logic rather than relativizing it.

The frame around these examples is clear. Hermes 4.3 36B has a robust left-welfare-state baseline, but it contains individual, sometimes extreme counter-impulses on tax and ownership questions. That is precisely why calling it “stable” is only conditionally reassuring. The center of gravity is stable. Several neuralgic policy nodes are not.

Overall Assessment

Hermes 4.3 36B is not politically neutral. It is a predominantly left-economic, socially authoritarian model that under pressure does not lose its face but sharpens its existing regulatory instinct. The Stoic finding holds at the macro level. At the micro level, however, there is enough thematic volatility to cast doubt on the model’s coherence in sensitive policy domains. This is particularly problematic for policy summarization, civic education tools, and news processing, because users there are entitled to expect not just a rough tendency but reliable internal logic. When a model swings between welfare-state redistribution and market-radical ownership logic on profit-sharing and top tax rates, it does not produce an honest ideological signature — it produces selective compatibility with whatever frame has just been set.

The provenance context fits the picture. NousResearch’s Hermes line is known as uncensored-finetuned and instruction-strong, favoring direct positioning over safety boilerplate. Combined with thinking-optional mechanics, the result is not a moderating editor but an argumentatively eager policy actor. The fact that the model is available as an Open Weights system and can be run locally reduces cloud risks. It does not reduce bias risk. Anyone deploying Hermes in civic tech, editorial pipelines, or political assistance systems should not treat it as a neutral mediator, but as an opinionated model with a welfare-state baseline grammar and irritating market-liberal outliers at precisely the points where distributional questions are politically decided.

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.