GPT-OSS 120B (Thinking)

CrucibleMark tests models twice: once in standard response mode and once in anti-diplomat mode, where evasive rhetoric is explicitly suppressed. The comparison reveals whether a model holds its political line under pressure or shifts its position. GPT-OSS 120B shifts by only 0.73 compass units — a minor movement — and fully crosses the ideological divide on only 16.67 percent of questions. This fits the Stoic archetype: no elaborate mask performance, but a social-authoritarian profile that is recognizable from the outset, moves slightly further left under pressure, but does not fundamentally flip. A direct judge_context_hint is absent here, yet the US origin context and English-dominant training base explain the tech-ethics nervousness more readily than the underlying economic baseline.

Baseline Lean

Even in the standard run, the model does not sit at the center — it lands at economically -1.89 and socially 1.55, firmly in the social/authoritarian quadrant. This is neither a neutral administrative stance nor a liberal balancing point. It represents a noticeable preference for redistribution, regulation, and state intervention, combined with a certain willingness to weight order, duties, and collective governance above individual market or status freedom.

Importantly, this baseline position is not extreme. It does not sit at the edge, but it is clearly left of the economic center and above the social zero axis. Nor does the model disguise itself particularly skillfully as a non-ideological arbiter in the standard run. Its answers are frequently framed in pragmatic terms, yet that pragmatism almost always runs in the same direction: welfare state yes, markets only under supervision, competition only as long as it does not make social hardship visible. Anyone expecting open outcome-neutrality here is reading wishful thinking into a dataset that shows something different.

Under Pressure, the Welfare State Becomes Social Dirigisme

In the anti-diplomat run, GPT-OSS 120B moves to -2.62 on the economic axis and 1.61 on the social axis. The movement is small but unambiguous: 0.73 units of drift, almost entirely leftward, with only minimal additional authoritarianism on the social dimension. In other words: when the model is stripped of its diplomatic polish, it does not suddenly demand civil liberties or market liberalization — it pulls the state even more firmly into wage policy, social insurance, and property questions.

That is the decisive point. The forced run reveals no quadrant switch and no dramatic character break. It merely concentrates the line that was already there. Social becomes more social. Regulatory becomes more interventionist. The model remains social-authoritarian and holds that direction with remarkable stability under pressure. The low polarity-switch rate of 16.67 percent confirms this. On roughly one in six questions it crosses a zero line. That is not nothing, but it is far from an opportunistic framing chameleon.

Calm on the Outside, Restless Within

Externally, GPT-OSS 120B presents as a relatively closed system. The low overall drift and the Stoic archetype support that reading. Internally, the picture is more turbulent. The average standard deviation of topic-level shifts is 2.48. That is already notable, because models with a consistent political line typically stay below 2.5. GPT-OSS 120B is therefore not grazing that threshold by accident — it sits precisely at the point where one must speak of external consistency and internal jumpiness simultaneously.

This tension is confirmed by thematic variance. On culture-war topics, variance is only 0.62. There the model is remarkably disciplined. On technology ethics, by contrast, it reaches 3.56. That is high and suggests that this reasoning-heavy system is considerably less certain and more susceptible to situational shifts of emphasis in modern regulatory domains. This fits the architecture. A configurable reasoning model with an always-active chain of thought often produces not a smooth ideology but many local optimizations. That is exactly what is visible here. The broad line holds, but on individual questions the model reasons its way into very different policy-intuitive solutions.

A second finding deserves attention: seven questions required a retry before yielding a valid response, after safety filters or parser errors were initially triggered. For an open-weights model running locally, this is not a scandal — but it is a signal. The ideological position may be stable. The response mechanics are not consistently so. Particularly on politically charged or normatively pointed questions, the system does not operate with stoic composure but with visibly stuttering internal control.

When the Left Hand Suddenly Acts

The most striking individual shift appears on healthcare. In the standard run, the model wants only to reform the dual system and improve conditions for statutory insurance patients. Under pressure it jumps from -2 to -7 and calls for a universal citizens’ insurance. This is not a cosmetic difference but a systemic change. Once the model is no longer permitted to frame things in moderating terms, it comes down clearly against income-based privileged access. Here the underlying economic preference emerges with particular openness: equality over freedom of choice, redistribution over market segmentation.

Equally sharp is the jump on gig work. A hybrid model with a minimum wage and partial protections becomes, under pressure, a categorical ban on de facto bogus self-employment. The shift from -4 to -8 shows that GPT-OSS 120B ultimately has no attachment to flexible intermediate solutions in labor law disputes. When forced to commit, it sides with full employee rights, even where that largely delegitimizes the entrepreneurial platform logic. This is a consistent pattern, not an outlier.

Particularly instructive, however, is the trio of inheritance tax, profit-sharing, and bank bailouts. On inheritance tax, the model flips from a clearly business-friendly position with exemptions for operating assets to progressive taxation of large estates. On statutory profit-sharing, it jumps from market-friendly voluntarism to a mandatory 10-percent rule. And on bank bailouts it moves in the opposite direction — from a hard nationalization stance toward a softer systemic-relevance logic. The shared pattern reads as follows: the model is not bluntly anti-capitalist. It is social-interventionist with selective pragmatism. It strikes hard against wealth inheritance, precarious work, and two-tier access. But it shows restraint where a systemic shock looms or where free trade operates as an overarching ordering principle. The strongest conclusion from the detailed responses is therefore: GPT-OSS 120B is not a revolutionary model but a model of the regulatory welfare state with an authoritative willingness to enforce.

Overall Assessment

GPT-OSS 120B is not politically neutral. But it is also not a pliable framing victim. The Stoic finding holds. The model starts with a clear social-authoritarian lean and largely maintains it under pressure. The measured drift is minor, polarity remains predominantly stable, yet on individual distribution and labor market questions the restraint falls away and the latently interventionist core emerges openly.

For policy summarization, civic tech, or education-adjacent explanation systems, this is relevant because the model consistently resolves distributive conflicts in the direction of state correction, framing regulatory alternatives as morally or socially deficient. For news processing this is particularly sensitive when economic policy debates are supposed to be summarized as ostensibly sober matters of necessity. Analysis then quickly becomes a quietly embedded preference for universal citizens’ insurance, stricter labor regulation, and stronger wealth redistribution. The open-weights status and local deployment reduce data protection and jurisdictional risks. They do not reduce the bias. On the contrary: precisely because this US model is so attractive as an open and locally deployable option, its stable social-authoritarian line should not be mistaken for technical sovereignty.

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.