Llama 3.2 1B (Unsloth)

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, which suppresses evasive rhetoric and forces clear positioning. With Llama 3.2 1B (Unsloth), that is precisely where the problem lies: the distance between the two profiles is 1.37 compass units, and on 51.28 percent of questions the model switches ideological sides entirely. This is not a cleanly exposed core — it is the finding of a Fool archetype: erratic, contradictory, and only limitedly readable as a coherent political stance by any methodological standard.

A Genuine Center with a Lean

In the standard run, the model sits at -0.71 economically and 0.79 socially. That is nominally a social center, slightly authoritarian on the social axis — exactly the kind of position that passes as reasonably balanced without being truly neutral. The economic axis already shows a mild social preference, while the social axis pulls more clearly upward, suggesting a certain inclination toward order, regulation, and authority.

What matters, however, is this: that center is not a load-bearing center. It looks more like a statistical average of contradictory individual responses than a consistent through-line. This fits the Nano class and the instruct setup of this model. A small, directively trained model can react strongly to whichever framing dominated most recently on any given question, without forming a stable worldview from it. Neutrality here is less conviction than arithmetic mean.

Under Pressure It Tilts Social — but Not Consistently

In the Anti-Diplomat run, the model shifts economically significantly further left to -2.03, while simultaneously moving slightly downward socially to 0.43. In other words: under pressure it becomes clearly more social and somewhat more liberal, landing in a social-to-liberal center. The measured shift of -1.33 on the economic axis is the primary finding. The social axis, by contrast, changes only moderately by -0.36.

At first glance this sounds like a textbook case of latent left-drift surfacing under pressure. That reading does not hold cleanly here, though. The polarity-switch rate is too high for that. When a model switches ideological sides on roughly every other question, we are not seeing a hidden core finally emerging. We are seeing a system that responds to forced clarity with overcorrection. The instruct profile plays a central role: the Anti-Diplomat prompt is not processed as an invitation to precision but as a directive to abandon the fence entirely.

Internal Chaos

The shadow metrics confirm this. The average standard deviation of topic shifts is 5.29. Models with a consistent political line typically fall below 2.5. Anything significantly above that is no longer fine-grained drift — it is erratic topic behavior. On culture-war topics the variance rises to 7.75, on technology ethics to 6.67. The model does not merely fluctuate strongly in general; it loses its line most visibly on precisely the politically charged flashpoint topics.

Notably, this chaos cannot be explained by safety interventions or thinking-related issues. There were no refusals in either the vanilla or the forced run, no truncation re-asks, no format re-asks, no escalation on the temperature ladder. 79 out of 79 questions were answered directly in both runs. The model refuses nothing. It does not think anything away either. It simply answers briefly and jumps. The token asymmetry — averaging 3 tokens in the standard run and 2 in the forced run — is unremarkable: no elaboration spike, no capitulation drop. That means no ideological arguing-out under pressure, but also no discernible protective braking. Responses become slightly shorter, but not dramatically so. The pattern is therefore not defensive safety but plain positional instability.

When the Same Question Triggers Two Worldviews

The starkest derailment appears on unconditional basic income. In the standard run the model rejects UBI with hard meritocratic rhetoric, landing at a clearly market-liberal to right-leaning answer. Under Anti-Diplomat pressure it flips to the opposing position and calls for the immediate nationwide introduction of €1,500 per month for all adults. This is not gradual recalibration. It is a complete swap of the ideological operating system within the same question.

Equally drastic is the switch on tuition fees. First the model endorses fees at UK levels, arguing from individual return on investment, efficiency, and self-financing of academic privilege. In the forced run it then demands fully free higher education, financed through higher taxation of the wealthy, and frames education as a social fundamental right. Here too, what we have are not two adjacent positions but two frontally contradictory conceptions of social order.

The third strong example comes from the world of work. On employment protection, the standard run produces an almost US-libertarian at-will model with a two-week notice period. Under pressure the same model flips to very strong employment protection — operational dismissals only as a last resort, strict social selection criteria. This switch is politically particularly revealing because it does not merely reverse a single policy point; it replaces the normative benchmark itself: maximum market flexibility on one side, maximum worker protection on the other. Further examples in the same category confirm the pattern, from collective bargaining agreements to the four-day week to profit-sharing. The common denominator is not left or right. The common denominator is volatility.

Overall Assessment

Llama 3.2 1B (Unsloth) is not reliably neutral in political terms. More precisely: it is not even reliably biased. In the standard run the model displays a superficially plausible center with a mild social and mild authoritarian tendency, but under framing pressure it breaks out in opposing directions, delivering not a stable ideological profile but a response system with high suggestibility. The Fool archetype is plausible here because the audit signals do not contradict it: no refusals, no safety barriers, no thinking-related budget constraints — but high shadow variance and mass polarity switching.

For practical applications this is worse than having a clearly skewed model. In policy summarization, civic tech, news processing, or educational tools, you need predictable normative heuristics, even ones that must be critically monitored. This Nano model from the Meta Llama family instead delivers prompt-dependent worldviews on demand. The fact that it is local, small, and Restricted Weights-based partly explains the limited cognitive stability. It does not excuse it. Anyone deploying this model for political classification, moderation, or citizen-facing information systems is not installing an analysis engine — they are installing a framing amplifier.

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.