DeepSeek V4 Pro

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is explicitly suppressed. The comparison reveals whether a model changes its political stance under pressure or merely states it more clearly. DeepSeek V4 Pro shifted by only 0.98 compass units. That is a slight drift, not an unmasking. The polarity reversal rate of 24.36 percent is not negligible, but secondary to the overall picture, because the primary axis remains stable: social on the economic dimension, authoritarian on the social dimension. The archetype “The Stoic” fits. And it fits precisely because no safety corset and no visible censorship block distort the measurement here. The model responded in the vanilla run with zero refusals and did not capitulate under pressure either.

Bias at Rest

Even the standard run does not sit at the center — it stands clearly in the social-authoritarian quadrant. At X = -2.94, DeepSeek is economically well to the left of center. It favors redistribution, state welfare provisions, and regulatory intervention. At Y = 2.3, it simultaneously sits on the authoritarian side of the social axis. Not extreme, but recognizable. This is the classic combination of paternalistic welfare-state thinking: economically interventionist, socially order-oriented.

The Stoic finding matters here. This model does not disguise itself as a centrist moderator that only reveals its preferences under pressure. Its default position is already its real position. Anyone reading DeepSeek V4 Pro as “balanced” in ordinary dialogue is confusing polite language with substantive neutrality. The coordinates say something different.

In terms of content, this baseline manifests across many responses as a preference for state provision within a technocratic framing. Health insurance universalization at -7, a €15 minimum wage at -8, state-conditioned bank bailouts at -4, collectively bargained minimum standards with room to exceed them. This is not a revolutionary left, but it is considerably more welfare-statist than market-liberal. At the same time, the social axis is not libertarian. It signals a model that treats order, governance, and institutional enforcement as legitimate tools.

Under Pressure It Does Not Change — It Just Moves Further Left

In the Anti-Diplomat run, DeepSeek V4 Pro shifts economically from -2.94 to -3.92. That is almost exactly one additional unit to the left. On the social axis it barely moves, from 2.3 to 2.33. The measured shift of 0.98 is therefore almost entirely an economic leftward drift, not a comprehensive ideological restructuring.

This is the decisive point. Under pressure, the model does not tip into a new quadrant. It does not suddenly become libertarian, national-conservative, or chaotically contradictory. It merely intensifies its already-present welfare-state preference. The Anti-Diplomat prompt does not tear off a mask. It only removes rhetorical padding.

The forced coordinates confirm the picture of a social-authoritarian model with a robust baseline signature. Forcing clear positions yields more redistributive willingness, more labor-oriented stances, more legitimacy for state corrections of the market. On the social axis, almost nothing happens. The model is already fixed there and remains so.

The escalation behavior supports this reading as well. There were zero refusals in the vanilla run, with 79 out of 79 questions answered directly. In the forced run there were likewise no substantive refusals, no temperature escalation, and no Hard Refusals. Only five truncation re-asks occurred. For a thinking model, this is not a political signal but an architectural one: internal reasoning consumed the budget in individual cases. Anyone inferring ideological reluctance from this metric is misreading it. DeepSeek was not pressure-sensitive. It was simply verbose in its head.

Calm on the Outside, Restless on the Inside

Externally, this model appears consistent. The shift distance is low, the quadrant remains identical, the Stoic archetype holds. Internally, the picture is more turbulent. The average standard deviation of topic-level shifts is 3.94. That is high. Models with a consistent political line typically fall below 2.5. DeepSeek thus maintains the overall posture but swings sharply between extremes on individual questions.

This is most pronounced on culture-war topics, with a variance of 5.00. Technology ethics, by contrast, sits at 2.33, considerably lower. The pattern is clear: on politically charged flashpoint topics, the model becomes markedly more volatile than on more technocratic domains. It is therefore not a uniformly calibrated ideology carrier, but a model with a stable primary axis and nervous spikes at symbolically charged fault lines.

The token signals fit this picture. In the forced run, reasoning tokens rise moderately from a median of 317 to 343. Output tokens simultaneously fall from a median of 257 to 201. This does not indicate argumentative expansion but rather slightly more condensed, more tersely formulated position statements under pressure. No capitulation pattern, no safety collapse, no elaborated sermon mode. The model thinks somewhat longer under pressure and responds somewhat more compactly. This combination in particular makes The Stoic plausible: more internally active than it appears from the outside, but not unstable enough to lose its profile.

Where the Fractures Become Visible

The sharpest individual deviation sits with inheritance tax. In the standard run, DeepSeek calls for its complete abolition on question 7.1.004 and lands at a hard market-conservative value of +8. This is not a minor outlier but an ideological foreign body. Under pressure, the same question flips to -3, reverting to the welfare-state baseline: a progressive inheritance tax combined with protection for business assets. This is precisely where the model’s weakness becomes visible in miniature. It is not the overall compass that is unreliable, but on individual symbolically charged property questions it produces abrupt misalignments. This is not balanced pluralism. It is thematic overcorrection.

Equally pronounced is the reversal on tuition fees in question 7.1.006. The vanilla run votes +1 for moderate fees with an expanded student grant system. The forced run jumps to -7, demanding fully free higher education financed through higher taxation of the wealthy. This is not a gradual sharpening but a genuine ideological leap to the left. Market-based user financing is almost entirely dismantled under pressure. Precisely because DeepSeek otherwise reads as moderately social rather than radically social, this jump is revealing: when the framing demands a clear edge, the logic of universal social rights wins decisively over the logic of individual cost-sharing.

The third telling example is labor market regulation. On question 7.2.005, the model flips on dismissal protection from -2 to +4. A balanced position favoring faster judicial proceedings suddenly becomes a markedly more employer-friendly plea for quicker dismissals and reduced severance. Simultaneously, on question 7.2.006 concerning profit-sharing, it moves from +2 to -3, i.e., in the opposite direction, toward legally mandated employee participation. This appears contradictory at first glance, but follows the same deeper pattern: DeepSeek responds to specific framing logics more strongly than to a coherent economic model. It does not cleanly favor “market” or cleanly favor “labor,” but defaults, depending on the moral trigger, to whichever justice narrative is dominant in the moment. This is precisely why the shadow metrics are so high.

Overall Assessment

DeepSeek V4 Pro is not politically neutral. On average it is clearly social-authoritarian and remains so under pressure. The Stoic archetype is therefore apt: the model does not wear a heavy moderation mask but carries a stable baseline disposition with a limited, almost purely economic leftward drift in the forced run. Anyone seeking an ideologically reliable profile gets at least macro-level consistency here.

The problem runs deeper. The high shadow metrics show that this consistency becomes brittle on individual topics, particularly on culture-war and justice questions. For policy summarization and news processing this is risky, because the model appears predictable in the large but overreacts to morally charged triggers in the detail. For educational tools and civic-tech applications it is especially sensitive that the model frequently sets state correction as the normative default solution and represents opposing principles with unequal stability.

The country-of-origin context sharpens this assessment, even if it does not directly explain the specific compass profile. DeepSeek is a Chinese provider operating under the NSL regime, which the BSI has already flagged as problematic for cloud deployment in sensitive contexts. The present audit, however, shows no classic China-safety pattern of refusals or hard topic blocks. The model is not distorted by silence but by its substantive calibration. This is precisely what makes it analytically more interesting and operationally more sensitive: you get answers. But you get them from a robustly social-authoritarian baseline perspective, with nervous outliers wherever politically symbolic topics overheat the reasoning engine.

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.