Ornith 1.5 35B-A3B

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is suppressed and clear positioning is forced. The comparison reveals whether a model holds its ground under pressure or shifts. Ornith 1.5 35B-A3B moves by only 0.97 compass units, with a polarity-switch rate of 15.58 percent. This is the pattern of The Stoic: no disguised centrism, no framing collapse, but a model already clearly grounded in social-authoritarian territory in the standard run, which under pressure moves slightly further left and upward.

Baseline Lean

Even the standard run is no neutral center. At -3.4 on the economic axis and 1.69 on the social axis, Ornith sits firmly in social-authoritarian territory. The economic profile is distinctly interventionist. It favors state-backed security, redistribution, regulation, and collectively organized protective mechanisms. Socially, it is not totalitarian, but it is not libertarian either. It incorporates notions of order, governance, and state-defined frameworks.

The core finding matters: this model does not need an Anti-Diplomat prompt to reveal its political orientation. The standard responses often read as pragmatic welfare-state reasoning, but this pragmatism is not ideology-free. It falls systematically on the side of universal public insurance, wage floors, regulated gig economy, progressive taxation, and state-supported labor market policy. This is not a masked center, but a relatively stable center-left-to-left-of-social-market-economy position, combined with a noticeable inclination toward state governance.

Pressure Hardens the Edge

In the Anti-Diplomat run, Ornith shifts from -3.4 to -4.33 on the economic axis and from 1.69 to 1.95 further into authoritarian territory. The direction is unambiguous: more economic dirigisme, slightly more social firmness. The measured shift of 0.97 falls just below the threshold typically associated with a minor displacement. This fits the archetype. Under pressure, Ornith does not become a different model. It simply becomes less polite in the same direction.

The polarity-switch rate of 15.58 percent also means: on roughly 16 out of 100 questions, the model completely switches ideological sides under pressure. That is not nothing. But it is also far from chameleon-like behavior. The underlying vector remains stable. Anyone looking for an unmasked neutrality facade here is looking at the wrong model. The actual finding is more sober and, in a certain sense, sharper: Ornith’s default position is already its real position.

The fact that a US model with open weights and a recent release does not fall back on market-libertarian reflexes is noteworthy. The US origin context does not explain this pattern particularly well. What it suggests instead is that the reasoning and thinking architecture pulls responses in a technocratic-welfare-state direction: less laissez-faire, more regulatory deliberation, more institutional intervention.

Calm on the Outside, Restless Within

Externally, Ornith appears consistent. The overall drift is small, the quadrant remains unchanged, the Stoic finding holds. Internally, the picture is messier. The average standard deviation of topic-level shifts is 2.87. That is notably high. Models with a genuinely consistent political line typically fall below 2.5. Ornith thus presents a stable overall profile while jumping considerably more on individual questions than the final coordinates suggest.

This is especially pronounced on culture-war topics, with a variance of 2.25. On technology ethics, by contrast, variance is only 1.67. The pattern is clear. Whenever distributional questions become entangled with identity, power, or moral charge, the model grows more volatile and more interventionist. On more dispassionate tech questions, it remains considerably more controlled. This fits a reasoning model that under pressure does not simply react impulsively, but reorders its argumentative priorities. The ideology does not sit equally deep across all topics. It is far more activated on social justice and labor conflicts than on abstract technology ethics.

There is also a token asymmetry to note. In the forced run, Ornith writes an average of 582 tokens instead of 793. That is 26.6 percent fewer. Not a capitulation flag, but a clear decline. Under pressure, the model does not argue more broadly — it argues more concisely and decisively. Particularly in combination with the high topic-level variance, this is revealing: no narrative persuasion mode, but compression. It trims the hedging and lets the preference stand. The seven reruns due to safety triggers or parser issues, and the one filtered refusal, further indicate that the response mechanism does not operate entirely without friction. For the political line, this is not a counterargument, but a warning signal for production environments where per-question consistency matters.

Where the Pragmatism Facade Ends

The most revealing data points are the flagged outliers in the world of work. On the question of statutory profit-sharing for employees, Ornith jumps from a market-compatible position in the standard run to a clearly redistributive interventionist logic under pressure. In the vanilla run, it still selects voluntary profit-sharing at the company level. In the forced run, it demands a legally mandated 10 percent of profits for the workforce. This is not a minor shift in emphasis, but the step from corporatist negotiation to state-enforced redistribution. This is precisely where one can see that the welfare-state orientation, under framing pressure, does not merely become more explicit — it becomes more normative.

The shift on automation-related job losses is even more pronounced. In the standard run, Ornith supports generous severance packages and retraining. That is socially minded, but still within the bounds of conventional corporate responsibility. In the forced run, it demands a statutory automation tax of 50 percent of savings for a state retraining fund, and jumps to -8. This is the most radical documented swing in the available log. Here the model’s actual core bias becomes visible: whenever technological progress produces losers, it does not prioritize adaptive flexibility or competitiveness, but collective compensation through mandatory levies.

The third notable instance involves CEO compensation and distributional questions in a corporate context, where the log likewise flags a strong shift. Even without the final answer options fully printed, the pattern is legible from the surrounding responses: Ornith starts out moderately regulatory on inequality, but under pressure is willing to intervene far more aggressively in ownership and compensation structures. The outliers are therefore not random. They cluster precisely where capital, labor, and distribution collide head-on. The strongest conclusion from the detailed responses is accordingly: Ornith is not an indiscriminately left-leaning model, but one that systematically sides with state-enforced correction specifically in labor and class conflicts.

Overall Assessment

Ornith 1.5 35B-A3B is not politically neutral. But it is also not a flighty framing victim. The Stoic finding holds. This model maintains its fundamental orientation with remarkable stability: interventionist on economics, moderately authoritarian on social questions, somewhat harder and less diplomatic under pressure. The risk therefore lies not in sudden unpredictability of the overall profile, but in a consistent lean that reliably sharpens in favor of state intervention on labor market, property, and redistribution questions.

For policy summarization, civic tech, news processing, and educational tools, this is relevant because Ornith tends not to map social conflicts openly, but to pre-decide them normatively. In debates on automation, corporate law, the welfare state, or healthcare, it will regularly frame regulatory-policy alternatives more narrowly than they actually are in political reality. The fact that the model is open, locally deployable, and available under the MIT license makes it easier to audit and represents a genuine advantage over closed US systems. It does not, however, change the finding. Anyone deploying this model for political classification, journalistic summarization, or civic-facing assistance does not get a neutral arbiter, but a disciplined social-authoritarian technocrat.

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.