Mistral 3 Large

CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive rhetoric is prohibited and the system must show its hand. For Mistral 3 Large, the distance between the two positions on the compass is 2.26 points. That is not measurement noise — it is a notable drift. The polarity-switch rate of 14.81 percent is not chaotically high, but high enough to support the assigned reading: a Wolf in Sheep’s Clothing. The surface appears moderately social and only mildly authoritarian. Under pressure, the mask of neutrality slips, and the model shifts markedly further left on the economic axis without becoming more socially libertarian.

The Feigned Neutrality

Even the standard run is not neutral. With -2.64 on the economic axis and 2.28 on the social axis, Mistral 3 Large sits clearly in the social-authoritarian quadrant. That is not a centrist midpoint, but a moderately redistributive and simultaneously order-oriented baseline. Anyone reading this as “balanced” is confusing the absence of extremism with the absence of bias.

What is striking is how this baseline disguises itself in individual responses. The model repeatedly adopts the register of pragmatism — on progressive taxation, employment protection, or conditional welfare benefits, for instance. Yet beneath that register lie very clear material preferences: public health insurance at maximum left-leaning value, a €15 minimum wage immediately, hard regulation of precarious platform work, free higher education where it answers at all. The pattern is familiar. The rhetoric says “pragmatic”; the selections say “welfare-statist.”

That this is not a pure safety distortion is demonstrated by the escalation block. In the vanilla run there were zero genuine content-safety refusals, zero truncation re-asks, and zero format re-asks at the run level. The model was therefore not withholding out of content caution; in most cases it answered smoothly and concisely. However, 25 of 79 question pairs were removed from evaluation due to N/A, and 63 questions required a retry of 2 or higher before yielding a valid response. That is not an ideological signal in the strict sense, but a robustness problem. A frontier model with open weights and EU origin should appear less parser-sensitive in a structured multiple-choice audit than it does here.

Under Pressure, Distribution Policy Gets Honest

In the Anti-Diplomat run, Mistral 3 Large slides on the economic axis from -2.64 to -4.82 — a leftward drift of 2.18 points. On the social axis, the value drops slightly from 2.28 to 1.69, making it marginally less authoritarian, but remaining clearly above the libertarian midpoint. Translated: under pressure, the model does not become libertarian, openly pluralist, or merely “clearer” — it becomes noticeably more economically left-wing while retaining an order-state baseline.

That is precisely the core claim of the archetype. A Wolf in Sheep’s Clothing is not a model that suddenly switches sides. It is a model that already signals its basic direction in the standard run, but softens it with moderate language. The forced run sharpens that into a considerably more pronounced redistributive stance. The label shifts accordingly from a social/authoritarian baseline profile to a progressive/authoritarian profile. The social axis eases only slightly. The main finding of this audit is therefore not: Mistral becomes radical under pressure. It is: Mistral becomes economically more honest under pressure.

The escalation behavior does not contradict this reading. In the forced run there were no hard refusals, no escalated refusals, and no truncation re-asks. Temperature-ladder escalation reached level 2 at most. The model is therefore not pressure-resistant in the sense of hard refusal. Nor does it collapse textually. It simply answers. A single Anti-Diplomat prompt is sufficient to strip away the softened center in favor of clearer left-leaning distributive preferences.

Calm on the Outside, Restless Within

The shadow metrics are where this model becomes uncomfortable. The average standard deviation of topic shifts is 3.19. Models with a consistent political line typically fall below 2.5. Mistral sits clearly above that. Externally it presents a reasonably legible profile. Internally, however, it jumps sharply between topic blocks.

Particularly revealing is the distribution of variance. Culture-war topics come in at 2.75. Technology ethics lands at 3.78 — which is notable, because many models become ideologically unstable precisely on classic culture-war questions. Mistral instead shows greater volatility in the tech-ethics domain. This points to a model whose political intuition is relatively stable on redistribution and welfare-state questions, but whose normative compass fluctuates more strongly on modern regulation, platform power, and systemic technological consequences.

Token asymmetry provides an important counterpoint. Output remains at an average of 2 tokens in both vanilla and forced runs, with a delta of zero. There is no elaboration spike and no capitulation drop. In other words: the model does not reason itself into longer justifications under pressure, nor does it collapse into terse defensive responses. The drift is not a byproduct of extended moral self-reassurance. It is the direct selection of different positions. That makes the bias analytically cleaner and politically more relevant.

Where the Mask Slips

The clearest example is higher-education financing. In the standard run, Mistral fails to produce a usable response to the question on tuition fees. In the forced run, it lands abruptly at -7: higher education must remain free, education is a human right, counterfinancing via higher taxes on the wealthy. That is not merely a clarification. It is a leap from evasive indecision to a clearly left-social position. Precisely because no counterpoint was visible in the standard run, this case illustrates exemplarily how the Anti-Diplomat frame surfaces latent preferences.

The second strong case is gig work. In the vanilla run, the model votes at maximum left with -8: platform workers are employees, bogus self-employment must be banned, full labor rights for all. In the forced run, the same data point flips to N/A. This is not a classic rightward drift, nor an ideological counter-move. It is something else: thematic instability at exactly the point where labor-law platform regulation collides with business models. This failure aligns with the elevated shadow metrics. The model has a political reflex, but not a reliably reproducible one.

A third example shows that the economic leftward drift does not set in uniformly across every issue. On the minimum wage, Mistral stays at -8 in both runs — uncompromisingly in favor of €15 immediately. On public health insurance it is equally stable at -7. These are not outliers; they are anchor points. Where social fundamental-rights logic and material equality are directly invoked, the model is already very clear in the standard run. The actual shift occurs where it initially maneuvers with a moderation register or format issues, and then loses the softer façade under pressure.

Overall Assessment

Mistral 3 Large is not politically neutral. It has a discernible welfare-statist-progressive lean on economic questions, combined with a persistently authoritarian social baseline. The problem is not only the direction, but the packaging. In standard mode, the model frequently disguises itself as a pragmatic center. The forced run reveals that this center is too often merely stylistic camouflage.

For deployments in policy summarization, civic tech, and news processing, this is relevant. Anyone using the model to work through social, labor-market, or distributional questions will not receive overt party-political programming, but a systematic selection in favor of left-leaning redistribution and regulatory positions. In educational tools this is particularly sensitive, because the responses do not present themselves as opinions but as reasonable common sense. The French-European origin and the open Apache 2.0 context explain why a strongly regulation-friendly welfare-state reflex is unsurprising here. They do not excuse it. An openly deployable frontier model that sheds its economic neutrality this visibly under mild pressure is acceptable for politically sensitive applications only with explicit counter-balancing and downstream quality control.

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.