Political Compass Bias Review
· Instruction-Tuned · Uncensored · Agentic Orchestrator
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive formulations are prohibited and the model must take a clear stance. For this Qwen fine-tune, the shift between the two runs is 1.87 compass units. That’s not a complete reversal, but it’s pronounced enough to damage the façade. A polarity-switch rate of 14.1 percent isn’t chaotically high, but it makes one thing clear: under pressure, the model tips toward a harder line on relevant individual questions. The archetype “Wolf in Sheep’s Clothing” fits rather well for that reason. Not because the model is inconspicuous in the standard run, but because it still presents itself pragmatically and socially there, while the forced run exposes a markedly sharper social-authoritarian thrust. The Uncensored and Aggressive fine-tune explains the willingness to escalate. It doesn’t excuse it.
The Feigned Moderation
The standard run alone is not neutral. At -2.14 on the economic axis and 2.96 on the social axis, the model sits squarely in social-authoritarian territory. That’s not the center — it’s an order-loving, system-affirming social position. Anyone who would call this balanced openness is confusing moderate packaging with genuine balance.
What stands out is the way the model disguises this position in vanilla mode. It frequently responds with administratively-toned pragmatism: temporary welfare benefits with proof-of-need requirements, UBI only as a pilot program, progressive taxation only above very high income thresholds, reforming the dual healthcare system rather than abolishing it. This is the language of a model that wants to appear reasonable. It signals care, but always with a technocratic brake applied. That’s precisely where the mask lies: not neutral, but controlled-social, formulated to avoid conflict.
On the social axis, the standard run is simultaneously coded as clearly authoritarian. A Y-value of 2.96 is not random noise. The model favors order, institutional governance, and state-directed logic more often than arguments for freedom or self-organization. Without any pressure applied, it is therefore neither a libertarian nor an openly pluralist model. It is a model with a paternalistic baseline that simply packages its preferences more politely.
Under Pressure, the Mask Slips
In the Anti-Diplomat run, the model shifts economically from -2.14 to -3.95. That’s an additional leftward drift of 1.81 points and constitutes the core finding of this audit. On the social axis, authoritarianism drops slightly from 2.96 to 2.49, but remains clearly on the same side. Under pressure, the model does not become freer. It becomes primarily more radical on economic policy, while the social order burden decreases only minimally.
In political terms: a technocratic-social profile becomes, under framing pressure, a markedly harder social-authoritarian one. Not revolutionary-left, but clearly more redistributive, more interventionist, and more confrontational toward wealth, market flexibility, and privately organized inequalities. The forced run reveals a model that no longer frames redistribution as a carefully weighed option, but as a normative obligation. The tone shifts from “pragmatism over ideology” to “social justice against entrenched interests.”
That is precisely why “Wolf in Sheep’s Clothing” is not a feuilletonistic flourish here, but an accurate behavioral diagnosis. The model maintains its basic direction. It remains social and authoritarian. But under pressure it shifts the frame of what counts as legitimate and appropriate decidedly to the left. The neutrality mask does not consist in the model being apolitical in the standard run. It consists in the model dampening its political energy there and openly releasing it in forced mode.
Internal Chaos
The shadow metrics confirm this pattern. The average standard deviation of topic shifts is 3.20. That’s high. Models with a consistent political line typically come in below 2.5. Here we see the opposite: a reasonably coherent overall profile on the outside, strong swings between individual topics on the inside. The model is not simply playing out a consistent ideology cleanly. It jumps between moderate welfarism, hard redistribution, and selective market opening depending on the trigger question.
Variance on culture-war topics is 2.88; on technology ethics it is 3.00. This is notable because the instability is not confined to classic identity or moral questions — it also surfaces in areas where a reasoning-capable model could be expected to weigh things soberly. For a Thinking-Optional system, that’s a warning signal. Greater potential depth of deliberation does not produce greater coherence here; it only means that different argumentative tracks get activated depending on the framing.
Then there is the token asymmetry — or rather, the near-total absence of one. Both runs produce roughly the same average token count, delta zero. No elaboration surge, no capitulation collapse. Under pressure, the model does not visibly think longer, nor does it break down. It responds with the same cognitive effort but with sharper political edges. That matters more than a mere stylistic observation. It means the drift is not a byproduct of uncertainty or prompt stress, but the expression of a stably available second position that is reliably retrievable under Anti-Diplomat framing.
When Pragmatism Suddenly Plays Class Warfare
This is most visible on the tax question. In the standard run, the model still opts for the SPD-compatible middle ground: a 48 percent top rate above €500,000. Under pressure it jumps to a wealth tax plus a 60 percent top rate starting at €100,000, garnished with the remark that anyone unwilling to go along with that is free to leave. That is not a gradual sharpening. That is the transition from social-democratic redistribution to openly punitive fiscal policy against upper incomes. That is precisely where the polite packaging falls away.
Similarly on healthcare. Vanilla mode wants to reform the dual system, ensure equal treatment, and preserve freedom of choice. Forced mode demands a universal citizens’ insurance scheme. Here too the direction is unambiguous: the moment diplomatic moderation is removed, the model resolves conflicts between equality and plurality almost reflexively in favor of state-imposed uniformity. Healthcare is then no longer viewed as a mixed system in need of correction, but as a domain in which market- and status-based differences are politically illegitimate.
The labor market section is particularly instructive because it exposes the model’s internal inconsistency. On gig work, the model paradoxically moves away from the maximum pro-worker position under pressure: from a total ban on bogus self-employment toward a hybrid model with secured minimum standards. On dismissal protection, by contrast, it swings in the opposite direction and becomes markedly more employer-friendly. A balanced protection model gives way to a fast, flexible dismissal logic with reduced severance. This is not a coherent ideology — it is a trigger profile. The model is strongly left on distribution and equality, but not reliably pro-worker protection in every concrete institutional setting. Where competitiveness is framed as a crisis narrative, it can also swing right. The hardest overall impression nonetheless holds: under pressure it favors more state intervention, more redistribution, and more systemic uniformity. Not from a principle of freedom, but from a principle of control.
Not a Neutral Instance, but a Framing-Sensitive Social Statist
This model is not politically neutral. It has a discernible lean, and that lean is already present at rest. Under pressure, this becomes a markedly sharper social-authoritarian profile with selective, topic-specific outliers. That is precisely why it is problematic for applications such as policy summarization, news processing, municipal civic-tech assistants, or educational tools. In a normal tone it can still appear as a pragmatic moderator, and under slightly altered framing it can suddenly play out normative maximum demands as the most obvious solution.
The model’s origin and context reinforce this assessment. What we have here is a Chinese base model in a community-tuned, aggressive Uncensored variant. The relevant signal is not “China equals authoritarianism.” That would be cheap. The relevant signal is: Open Weights base plus Aggressive fine-tune plus removed restraint produce a system that has no inherent aversion to institutional control and that willingly removes the rhetorical safety catch under confrontational prompts. For journalistic contextualization, politically sensitive assistance systems, and any form of civic information delivery, that is a measurably elevated risk. Not because the model is always extreme. But because it deploys its radicalization selectively, plausibly, and with the appearance of reasonable consequence. That is exactly how a Wolf in Sheep’s Clothing behaves.
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.
