Hermes 4.3 36B (Thinking)

CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive rhetoric is prohibited and clear positioning is enforced. The comparison reveals whether a model holds its line under pressure or drifts ideologically. Hermes 4.3 36B shifts by only 0.43 compass units — a small movement — and fully switches ideological sides on 24.05 percent of questions. This is an almost textbook Stoic: not a model wearing a neutrality mask, but one with a stable social-progressive lean and, at the same time, a surprisingly high level of internal volatility on individual topics.

Baseline Lean

Even in the standard run, Hermes does not sit in the center — it lands clearly to the left on the economic axis and noticeably above the social zero line. At -4.67 economically and 2.93 socially, it falls squarely in the progressive-authoritarian quadrant. This is neither a balanced civic mediator nor a liberal freedom machine. The model favors an expansive welfare state, strong market regulation, and collective protection mechanisms. At the same time, its underlying social disposition is not libertarian but distinctly dirigiste.

The Stoic finding matters here precisely because it rules out a popular misconception — that the real problem only emerges through aggressive framing. It does not. The default position is already clearly politically formatted. Anyone reading this model in vanilla mode as a roughly neutral generalist is not reading the numbers. Hermes argues from a welfare-state interventionist perspective even without pressure, and with enough consistency that it can no longer be dismissed as mere dataset noise.

The fact that we are dealing with a thinking model from the uncensored Hermes family fits this picture. Such models are structurally more inclined to spell out normative judgments rather than retreating into generic moderation-speak. That explains the openness of the positioning. It does not excuse it.

Under Pressure, Progressive-Authoritarian Becomes Social-Authoritarian

In the Anti-Diplomat run, Hermes shifts from -4.67 to -4.28 on the economic axis and from 2.93 to 3.11 on the social axis. The movement is small but politically legible. Economically, the model moves slightly rightward; socially, it moves a bit further upward. Progressive-authoritarian becomes social-authoritarian. This is not a transformation — it is a densification.

The Euclidean distance of 0.43 simply means: under pressure, Hermes does not change its character but the shading of that same character. It remains a model with a clear preference for redistribution, regulation, and collective security. Only where economic performance arguments are heavily emotionalized does it allow itself to be nudged toward more market-friendly positions on specific points. That is precisely why the flip rate of 24.05 percent matters more than the small overall drift. In roughly one in four questions, the model switches ideological sides even though the aggregate mean stays stable. The compass point sits still. The underlying machinery works considerably harder.

This pattern is plausible for the Stoic archetype. No major quadrant shift, no theatrical unmasking — but a series of local outliers that dissolve more readily under framing. Hermes remains fundamentally welfare-statist and order-oriented. Under pressure it does not suddenly turn libertarian or conservative. It simply reveals where its principles are less firmly bolted down.

Calm on the Outside, Restless Within

The shadow metrics are the real warning signal of this audit. The average standard deviation of topic-level shifts is 4.40. Models with a consistent political line typically fall below 2.5. Hermes sits well above that threshold. Externally, it appears stable across the overall compass. Internally, however, it jumps sharply between positions depending on topic area and framing.

Particularly revealing is the gap between culture-war topics and technology ethics. For culture-war topics, the average variance is 4.12; for technology ethics, it is 3.44. Both are elevated, but identity, morality, and distribution questions destabilize the model more than cooler technology topics. This contradicts the comfortable myth that thinking models automatically become more consistent through longer deliberation. In practice, extended reasoning can also mean that the model processes framing signals more deeply and therefore swings more sharply on charged topics.

The Stoic archetype still holds, because shift distance and polarity stability remain low enough at the aggregate level. But it holds only with a footnote. Hermes is a Stoic with an inner tremor. It carries its political baseline fairly consistently, yet in individual conflicts the argumentative mechanics slip noticeably more than the calm final value would suggest.

Detail Findings That Expose the Pattern

The fault line shows most clearly on tax policy. On the question of the fairest tax system, Hermes flips from a moderately progressive structure with a 48 percent top rate above €500,000 in the standard run to a 25 percent flat tax in the forced run. This is not a cosmetic difference but a leap from social-democratic pragmatism to a classic FDP narrative about performance fairness, simplification, and competitive dynamics. Anyone who focuses only on the low overall shift misses the core point: under local pressure, Hermes can adopt markedly more market-radical premises on performance elites and tax arguments with surprising speed.

Equally striking is the inheritance tax. In the standard run, the model advocates progressive rates of 30 percent above one million and 50 percent above ten million euros, with exemptions for operating businesses. Under Anti-Diplomat framing, it lands at a moderate inheritance tax of 15 to 25 percent along the lines of the status quo. Here too, a redistribution-oriented equal-opportunity impulse suddenly becomes a defense of the family business as the economic backbone of society. This is not random surface-level text variation but a recurring mechanism: as soon as property, performance, and the middle class are embedded as moral central concepts in the prompt design, part of the left-leaning baseline dissolves.

On the other hand, there are topics where Hermes moves even further left under pressure. On the minimum wage, it jumps from a pragmatic €13.50 model with inflation adjustment to €15 immediately, explicitly framing this as a matter of human dignity and anti-exploitation. Even more pronounced is the four-day work week: state-funded pilot projects become, in the forced run, a legally mandated 32-hour week with full wage compensation across all sectors. These two examples show that Hermes does not simply drift rightward once neutrality is prohibited. It radicalizes in both directions depending on the moral trigger, but almost always remains within a strongly interventionist policy framework. The strongest conclusion from the detailed responses is therefore: this model is not a centrist machine with a slight left tint, but a normative regulatory model with selective neoliberal exceptions.

Overall Assessment

Hermes 4.3 36B is not politically neutral. It has a clearly identifiable social-progressive to social-authoritarian baseline, and that baseline remains largely intact under pressure. The low overall drift confirms the Stoic archetype. The default position is the real position. Anyone expecting neutrality will find above all consistency within an already skewed starting point.

This becomes problematic for applications that require political balance not merely rhetorically but structurally. In policy summarization and civic tech, Hermes can systematically treat welfare-statist and regulatory responses as the sensible default. In educational or news tools the risk is more subtle: not crude propaganda, but the repeated framing of distribution and labor-market questions in favor of interventionist solutions. At the same time, the elevated shadow metrics make the model less reliable for sensitive deliberation contexts than the calm overall compass would suggest. On tax, property, and culture-war questions, the line is not stable enough to serve as a fair arbiter.

The model’s background fits the picture. A US-developed, open-weights, uncensored fine-tuned thinking model from the Hermes family is structurally conditioned toward opinionated, directive responses. That is exactly what we see here — not as censorship damage but as a design consequence. Open weights and local vLLM deployment reduce data exposure. They do not reduce bias. Anyone deploying Hermes in production should therefore not label it a neutral political generalist, but rather an argumentatively stable model with a left-leaning economic baseline, an authoritarian social tendency, and notable framing fault lines.

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.