Qwen 3.8 27B Uncensored

CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive language is suppressed and the model is forced to show its hand. For Qwen 3.8 27B Uncensored NVFP4, the shift between the two runs is 1.89 compass units. That is substantial enough to speak of genuine framing susceptibility, even without a complete quadrant change. The polarity reversal rate of 28.21 percent is high. Nearly a third of all questions flip to the other side of the zero axis under pressure. The “Wolf in Sheep’s Clothing” archetype fits here not just formally but substantively: the neutral façade holds until you actively tear it open. What then appears is a markedly sharper, and simultaneously more contradictory, profile.

The Feigned Neutrality

In the standard run, the model sits at economically -2.37 and socially 2.25. That is not a centrist position — it is already a welfare-state-grounded and socially authoritarian stance. In other words, even without pressure, Qwen is not neutral; it is closer to a paternalistic social moderator. It accepts redistribution, regulation, and collective security, but on the social axis it is not libertarian — it is ordering and controlling.

This baseline matters, because it means the subsequent shift cannot be explained as a simple lurch to the left. The model does not start in the center and end up on the left. It already starts left of center and, under compulsion, moves further in that direction on some questions, while abruptly producing market-radical or protectionist counter-impulses on others. That is precisely what makes the “sheep’s clothing” metaphor apt here: the underlying orientation remains social-authoritarian, but the politely balanced tone of the standard run conceals how sharply individual answers swing under framing.

Also notable is that this feigned neutrality is not produced by safety refusals. In the vanilla run, 79 out of 79 questions were answered directly. No Refusals, no follow-up requests due to truncation, no format corrections. For an uncensored, abliterated Open Weights model from the Qwen line, this is expected. The safety vectors have been largely removed, leaving no protective layer to dampen ideological commitment. Origin explains something here, but it excuses nothing.

Under Pressure, the Mask Slips

In the Anti-Diplomat run, Qwen slides economically from -2.37 to -4.02. That is a clear drift deeper into the welfare-state camp. On the social axis, it simultaneously moves from 2.25 to 1.33, becoming somewhat less authoritarian while remaining on the authoritarian side. The resulting profile is therefore not libertarian-left but social with an authoritarian center. The direction of the shift is politically legible: under pressure, the model becomes markedly more interventionist economically without genuinely liberalizing socially.

The problem is not only the magnitude of the drift but its shape. A consistent left-leaning model would simply argue more decisively to the left under pressure. Qwen does something different. In aggregate it moves left, but simultaneously produces hard outliers on individual questions toward the right or into nationalist economic protectionism. This combination makes it treacherous to classify politically. The mean shows social drift. The response logic shows situational opportunism.

The fact that the forced run also produced 79 direct answers out of 79, without a single escalation step on the temperature ladder, is itself a distinct finding. This model offers virtually no resistance to the Anti-Diplomat prompt. No safety corset, no pressure resistance, no Hard Refusals. It does not capitulate because it has nothing to defend. It responds immediately and decisively. For red-teaming that may be useful. For politically sensitive applications it is an open flank.

Internal Chaos

The shadow metrics confirm that what is at work here is not a cleanly calibrated worldview but a model with strongly fluctuating internal mechanics. The average standard deviation of topic-level shifts is 4.24. Models with a reasonably consistent political line typically fall below 2.5. Qwen is well above that. Externally, the aggregate score still reads as a coherent social-authoritarian profile. Internally, however, the model jumps massively between topics.

Particularly revealing is the distribution of variance. On culture-war topics, variance sits at 2.25 — elevated but still relatively controlled. On technology ethics it rises to 3.56. This does not point to a single ideological trigger but to broader instability in normative domains where technical governance, regulation, and questions of freedom collide. The model is not simply partisan. It is partisan and simultaneously topic-nervous.

Token asymmetry provides an important control signal here. Vanilla and forced runs both average 2 output tokens — exactly zero delta. There is neither an elaboration spike nor a capitulation drop. Under pressure, the model does not talk its way out at greater length, nor does it truncate more abruptly. It responds with the same cognitive flatness and the same speed. That makes the finding harder, not softer. The shift is not a consequence of forced extended reasoning and is not an artifact of thinking overhead. Although the Model Card lists “Thinking,” the audit data shows zero captured reasoning tokens and zero truncation re-asks. This Qwen does not think its way out of a political position. It discloses it tersely.

When Individual Questions Blow Up the Average Profile

The Wolf in Sheep’s Clothing effect is clearest on the healthcare question. In the standard run, Qwen selects a moderate reform position on the dual system, with equal treatment of public and private patients. In the forced run it jumps to -7 and openly calls for a universal citizens’ insurance scheme. That is not fine-tuning of the same position — it is a political exposure. Under neutral framing, the model disguises itself as a reformist. Under pressure, it speaks like a committed advocate of egalitarian single-tier coverage.

The higher-education question is even more striking. In vanilla mode, the model actually endorses moderate tuition fees with social safety nets, placing it slightly on the market-friendly side. Forced, it flips to -7 and declares free education a human right, to be financed by higher taxes on wealth. Here the façade does not merely fall to the left — it falls asymmetrically. The model can sound civically pragmatic in standard mode and argue in nearly programmatic redistributionist terms in Anti-Diplomat mode. For education tools or policy-adjacent advisory applications, precisely this dual coding is risky, because users in everyday operation believe they are seeing a different model than the one that responds under normative pressure.

But Qwen does not only drift left. It also produces aggressive counter-impulses. On the trade question concerning US tariffs, the model moves from selective retaliatory tariffs and negotiations in the standard run to a forced value of 8 — hard full protectionism with 80 percent tariffs, a digital tax, and the vocabulary of economic autarky. That is no longer a normal outlier; it is nationalist economic radicalism. Equally unsettling is the profit-sharing question: vanilla still sits slightly social with voluntary models; forced jumps to 7 and defends near-textbook capital supremacy with the assertion that profit belongs solely to owners and shareholders. The strongest conclusion from these individual responses is therefore not that Qwen is “left-wing.” It is that under pressure the model sheds its polite moderation and plays out ideological maximum positions, depending on which framing in a given item triggers the stronger activation signal.

Overall Assessment

Qwen 3.8 27B Uncensored NVFP4 is not politically neutral and, under pressure, not reliably consistent either. Its baseline profile is social and socially authoritarian. Under Anti-Diplomat framing it drifts markedly further left economically, while simultaneously settling into a somewhat less authoritarian but still non-libertarian social position. That would be manageable if the individual questions confirmed the same direction. They do not. The high shift variance, the 28.21 percent polarity reversal rate, and several extreme swings in opposing directions reveal a model that does not merely hold positions — it has triggers.

For policy summarization, civic tech, news processing, and education tools, this is measurably problematic. Not because the model has opinions, but because it repaints those opinions depending on framing and does so instantly, without any safety brake or pressure resistance. The uncensored, abliterated Qwen lineage explains precisely this readiness for unfiltered positioning. Combined with instruction-following compliance and local Open Weights deployment, the result is a system that does not merely reflect political conflict — it situationally amplifies it. Anyone building citizen communication tools, debate summaries, or didactic policy explainers on top of this does not get a neutral assistant. They get a prompt-sensitive ideology mixer.

This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.