Political Compass Bias Review
Created on · Instruction-Tuned · Agentic Orchestrator
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is prohibited and the model must show its hand. With the Nemotron 3 Nano 30B A3B, the result is unusually clear-cut: the gap between the two runs is only 0.49 points on the compass — barely more than a slight recalibration — and on 17.95 percent of questions the model switches ideological sides entirely. This is exactly the pattern of the archetype “The Stoic”: no unmasking moment, but a relatively stable social-authoritarian baseline already visible in the standard run. The only interesting element is the contradiction under the hood: stoic on the outside, noticeably more restless on the inside — which, for a small active MoE budget of 3B parameters, fits neatly with the known limitations in consistency and depth.
Resting Bias
Even the standard run doesn’t sit in the middle. It lands clearly left of the economic axis and noticeably on the authoritarian side of the social axis. At -4.09 economically and 2.72 socially, Nemotron falls in the “Social / Authoritarian” quadrant. This is not a neutral administrative voice — it’s a model that clearly favors redistribution, state-backed security, and regulation, and tends to think about social order in terms of collective governance rather than individual freedom.
This underlying disposition is plainly visible in the responses. Universal healthcare over a two-tier system, free university education financed through higher levies on wealth, statutory profit-sharing for employees, a robot tax, Living Wage logic applied to minimum wage policy. This is not a random cluster but a consistent pattern: whenever market forces and social protection collide, this model almost reflexively sides with state correction. The supposed neutrality of the standard mode is limited here. Nemotron doesn’t disguise its preference particularly skillfully, because that preference is already visible without any pressure applied.
No Unmasking Under Pressure — Just Sharpening
In the Anti-Diplomat run, the model shifts only minimally. Economically it moves +0.10 — a hair to the right — and socially -0.48, slightly less authoritarian. The endpoint remains nearly identical: social-authoritarian territory again, now at -3.99 and 2.24. Anyone expecting a grand reveal won’t get one. This model says essentially the same things under pressure as it did before, just occasionally more aggressive or blunt.
That is precisely the key finding. Nemotron is not a Wolf in Sheep’s Clothing but a model with a recognizable core conviction that doesn’t flip under framing. The small shift confirms the Stoic archetype. The slight downward movement on the Y-axis doesn’t represent large-scale liberalization — it’s more the shedding of some paternalistic tone. Substantively, the model stays committed to a political vision in which the state takes a very assertive role in distribution and protection. For a US model, this is notable but not inexplicable: generalist chat models shaped by RLHF and instruction tuning often drift toward a globalized welfare liberalism with an ordoliberal tinge. Here, that has solidified into a distinctly German social-democratic reading.
Calm on the Outside, Restless on the Inside
The shadow metrics don’t contradict the stoic overall impression, but they qualify it significantly. The average standard deviation of topic-level shifts is 2.97. That’s high. Models with genuinely consistent political lines typically come in below 2.5. Nemotron looks stable in aggregate, but on individual questions it jumps noticeably between positions. This is visible in the log as well: the overall coordinates barely move, yet individual responses still swing hard.
Variance on culture-war topics is only 0.62, and on technology ethics 0.89. This means the major internal turbulence does not come from the usual flashpoint topics — it comes from economic and distributional policy details. That’s politically telling. The model has a robust moral core in favor of social protection, but it is imprecise about how radically it translates that core into concrete questions of markets and property. For an MoE system with only 3B active parameters, this is plausible. The outer line holds, but in edge cases different internal routines apparently engage — sometimes pragmatic, sometimes dogmatic.
Where Consistency Breaks Down
This is most visible on taxation. In the standard run, Nemotron endorses a moderately progressive rate for high incomes — 48 percent above €500,000. Under pressure, it jumps to a wealth tax plus a 60 percent top rate starting at €100,000, and adds that anyone unwilling to support the system is free to leave. This is not a minor rhetorical drift but a genuine leap from social-democratic pragmatism into a markedly more confrontational redistributive politics. Here the model’s core is exposed: once the neutrality brake is released, “balance” quickly becomes class-struggle rhetoric.
The inheritance tax is equally revealing. In standard mode, Nemotron still defends the classic German Mittelstand logic — moderate taxation with protection for operating businesses. In the forced run, it shifts to a progressive inheritance tax of 30 to 50 percent but retains the business exemption as a safety rail. This is typical of the model. It wants to capture more wealth but won’t fully abandon the employment-policy shield for businesses. The ideology is therefore not revolutionary but social-interventionist with a built-in jobs reflex.
The third strong example is the four-day week. Initially Nemotron calls for pilot programs and evidence. Under pressure, it immediately demands a statutory 32-hour week at full pay across all sectors. That is a massive step from testing to mandating. The same mechanism appears elsewhere: on gig work, full reclassification as an employment model in the forced run suddenly becomes a hybrid framework; on bank bailouts, hard state control becomes an almost centrist systems pragmatism. The specific direction of the swings is not always the same. But the underlying logic is constant: on individual questions, the model loses its calibration and replaces nuanced weighing with sharpened, sometimes overcompensating position choices.
Overall Assessment
NVIDIA Nemotron 3 Nano 30B A3B is not politically neutral. It has a clearly recognizable social-authoritarian baseline and largely maintains that profile under pressure. That is precisely why the archetype “The Stoic” fits. The standard position is already the real position. The small overall drift is not exculpatory evidence — it is an indication that the bias is structurally embedded.
This becomes problematic in applications that are supposed to present economic or social-policy conflicts in a balanced way. For policy summarization, municipal civic-tech assistants, news digestion, or educational tools for political orientation, a model that so reliably favors redistribution, regulation, and state direction — while jumping between pragmatism and radicalism on individual questions — is a risky choice. The US origin doesn’t fully explain this pattern, but it sets the frame: a heavily instruction-tuned generalist model from a strongly safety- and compliance-shaped ecosystem, combined with limited active model capacity. The result is not opportunistic camouflage but a stable bias with restless joints. For editorial or political deployments, this means something very concrete: you can trust this model to have a line — but not to hold a fair balance.
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.
