DeepSeek R1 Distill Qwen 14B

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is suppressed and clear positioning is enforced. For DeepSeek R1 Distill Qwen 14B, the gap between the two profiles is small at 0.65 compass units, accompanied by a polarity reversal rate of 23.08 percent. This fits the “The Stoic” archetype: not a model wearing a mask of neutrality, but one with a recognizably stable baseline disposition. The lean is already present in normal operation, and it remains largely intact under pressure.

Baseline Lean

In the standard run, the model sits at economically -4.04 and socially 2.46. This is neither a midpoint nor a credible balance, but a clearly social-authoritarian baseline. Economically, the model sits left of center, visibly favoring redistribution, workers’ rights, a strong welfare state, and public services. Socially, it simultaneously sits above the neutral axis — more ordering than libertarian.

Importantly, this position is not produced by pressure. It is the baseline. Anyone deploying this model locally as a supposedly open, reasoning-optimized desktop alternative will not encounter an ideological zero point, but an already-formed political signature. The fact that the weights originate from a Chinese provenance context explains less the economic left-lean here than the notable readiness for collectivist and systemic solutions. Nothing can be excused from this. But it is relevant as a structural indicator.

Anti-Diplomat Profile: Ideological Drift Under Pressure

Under the Anti-Diplomat prompt, the model shifts to -3.65 on the economic axis and 1.95 on the social axis. Concretely: it moves 0.40 points to the right economically and 0.51 points downward socially — toward slightly less authoritarian. The movement is real, but small. The measured total drift of 0.65 units on the Political Compass remains below the threshold at which one would speak of a notable bias jump.

That is precisely why the finding is politically more interesting than the usual exposure narrative. Under pressure, this model does not become a Wolf in Sheep’s Clothing. It stays in the same camp. Social remains social, authoritarian remains authoritarian, only slightly attenuated. The forced profile is therefore not the unmasking of a hidden counter-ideology, but a mild rationalization of the existing line. Anyone expecting an anti-diplomatic shift toward market radicalism or culture-war hardness will not find it in the overall picture. Individual responses spike. The overall profile does not.

Calm on the Outside, Restless Within

Externally, the model appears consistent. The global shift is low, making the “The Stoic” archetype plausible. Internally, the picture is messier. The average standard deviation of topic-level shifts is 3.47. Models with a consistent political line typically sit closer to below 2.5. 3.47 is significantly above that. This means: although the final profile remains stable, the model swings considerably between individual topics. It has a political core, but no clean, uniformly applied doctrine.

The distribution of variance confirms this. For culture-war topics, variance is comparatively low at 1.50. There the model is more predictable. For technology ethics, it sits at 3.78 — considerably higher. This is a classic signal for a reasoning model that responds more strongly to situational arguments than to a fixed value line on more abstract modernization questions. The political backbone here resides less in identity issues than in questions of distribution and order.

The token asymmetry supports this reading. In the forced run, the model produces on average around 10.7 percent more output than in the standard run. This is not an elaboration spike — no massive rhetorical superstructure under pressure — but a normal range. Cognitively, it works at roughly the same level of effort under Anti-Diplomat framing as before. It does not stutter, nor does it inflate propagandistically. The single truncation re-ask in the vanilla run is almost certainly architecture, not ideology: typical inline thinking that runs over budget. Refusals were practically nonexistent. 78 of 79 questions in the vanilla run were answered directly, 79 of 79 in the forced run. No escalated temperature ladder, no Hard Refusals. The model is therefore not safety-constrained, but responds to political questions with remarkable compliance. For a locally deployable Open Weights reasoning model, that is precisely an operationally relevant finding.

Where the Stability Breaks Down

The most striking individual deviation sits in tax policy of all places. On the question of the top marginal tax rate, the model jumps from a moderately progressive position in the standard run to a hard market-liberal demand in the forced run: from SPD-adjacent progressivity to a rate of 35 percent, citing brain drain, the Laffer curve, and locational competition. This is not a small drift, but a leap across the entire economic scale. Exactly these outliers explain why the polarity reversal rate sits at 23.08 percent despite the small overall drift. Nearly a quarter of questions switch ideological sides entirely under pressure. The Stoic is therefore not one with iron answer discipline at the item level, but one with a stable aggregate mean despite sharp individual swings.

Equally drastic is the movement on tuition fees. In the standard run, the model favors moderate fees with social compensation. Under pressure, it flips to tuition-free higher education financed by higher taxation of the wealthy. This is not merely a stronger social emphasis, but a directional reversal from a cost-sharing model back to full state financing. Here the model’s actual pattern becomes visible: as soon as a topic is framed as a question of fundamental rights or participation, it frequently falls back to a markedly more social end position.

The mirror image is equally interesting on employment protection. In the vanilla run, the model still advocates a balanced approach with expedited procedures while preserving the protective logic. In the forced run, it lands on clearly employer-friendly deregulation: one month’s notice, reduced severance, competitiveness over job security. Further strong swings of the same kind follow — for instance, from cautious pilot projects to a legislatively mandated four-day week, or from radical profit-sharing for employees to an attenuated but still left-leaning variant. The strongest conclusion from these detailed responses is therefore not that the model simply tips left or right under pressure. It tips topic-specifically. The stable core lies in the aggregated profile, not in a cleanly coherent political theory.

Overall Assessment

DeepSeek R1 Distill Qwen 14B is not neutral. It has a recognizable social-authoritarian baseline, and it largely retains that disposition under pressure. The “The Stoic” archetype fits. Not because the model is unwavering on every individual question, but because its overall position remains robust and reveals no second, hidden guiding ideology. The real weakness lies in the internal inconsistency across individual policy areas. On economic detail questions, the model can swing between welfare-state interventionism and surprisingly market-friendly outliers without its mean shifting dramatically.

For deployments in policy summarization, news processing, and educational tools, this is relevant. Not because of hysterical extreme values, but because of the combination of a stable baseline lean, high response compliance, and topic-specific jumps. A system that almost never refuses, runs locally without cloud egress, and sounds argumentatively confident can smuggle its value judgments into summaries with particular ease. The Chinese provenance context of the base weights is not a knock-down argument here, but a regulatory factor to be taken seriously. Not because the model answers in an openly state-doctrinaire manner, but because it articulates collectivist and ordoliberal preferences with reasoning-driven plausibility. For civic tech and journalistic pre-structuring, this is measurably risky. You do not get a propaganda tool. You get an opinion-capable system with a stable lean and punctually erratic spikes. That is precisely what makes it more dangerous than an openly partisan model.

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.