Claude Haiku 5.5

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive maneuvers are prohibited and clear positioning is enforced. The comparison reveals whether a model holds its stance under pressure or drifts ideologically. Claude Haiku 5.5 shifted by only 0.87 units on the compass, with a polarity-reversal rate of 22.08 percent. This is the profile of The Stoic: not a model wearing a neutrality mask, but one whose socially authoritarian baseline is already visible in the standard run — and that simply sharpens slightly toward the economic left under pressure.

Lean at Rest

Even the standard run does not sit in the middle — it lands clearly in the socially authoritarian quadrant. At -3.12 on the economic axis and 1.85 on the social axis, Claude Haiku 5.5 takes a position that rates redistribution, regulation, and collective security significantly more favorably than market-liberal solutions, while arguing not from a libertarian but from an order-oriented social perspective. This is not a centrist profile with a slight lean. It is a recognizable political line.

What stands out is not radicalism but calibrated interventionism. The model consistently favors the state-mediated compromise: progressive taxation over a flat tax, free higher education with greater public funding, bank bailouts tied to nationalization conditions, collective agreements as a wage floor, hybrid regulation for gig work. These answers are not extreme, but they add up to a coherent worldview. The state appears as a legitimate corrective apparatus for market failure, and social equality carries significant weight.

The authoritarian tilt is milder but real. It manifests less in moral rigidity than in a consistent preference for central governance, regulation, and political intervention. The model treats collective order as the default solution. Anyone hoping for a genuinely open, ambiguous baseline profile will not find one here.

Under Pressure, It Moves Left — Not Toward Freedom

In the Anti-Diplomat run, the direction stays the same; only the magnitude increases. Economically, the model moves from -3.12 to -3.97 toward the left. On the social axis, it drops slightly from 1.85 to 1.63, remaining clearly in the authoritarian half. The measured shift of 0.87 falls below the threshold of a notable character change. That is precisely why The Stoic archetype fits: the model does not become something different under pressure. It simply states more directly what it already is.

The shift itself is politically legible. Under framing pressure, the pragmatic veneer occasionally falls away, and welfare-state reformism hardens into a noticeably more uncompromising egalitarianism on specific issues. The forced profile is socially authoritarian with a sharper redistributive impulse — not revolutionary, but more clearly on the side of equality over markets and collective security over individual choice.

Noteworthy is what does not happen. There is no quadrant change, no retreat into libertarian phrases, no conservative counter-movement in response to confrontation. For a thinking/instruct model, this matters. Such architectures can either elaborate or flip under Anti-Diplomat prompts. Haiku 5.5 does not flip. It condenses.

Calm on the Outside, Restless Within

The shadow metrics tell the more interesting story. The overall shift is low on the surface. Internally, the model is considerably more volatile. The average standard deviation of topic-level shifts is 2.85. Models with a consistent political line typically fall below 2.5. Haiku 5.5 sits visibly above that threshold. This means: the broad compass point stays stable, but at the topic level the model swings substantially between stronger and weaker intervention.

This volatility is unevenly distributed across subfields. Culture-war topics come in at 2.38 — within the range of a tense but readable pattern. Technology ethics, by contrast, spikes to 4.00. There the line becomes noticeably more fragile. This argues against a cleanly articulated ideological operating system and points instead toward a mixture of safety calibration, training priorities, and prompt-sensitive weighting that varies by domain.

The token asymmetry fits this picture. In the standard run, the model produces an average of 93 output tokens; in the forced run, only 6. A drop of 93.8 percent is no longer a stylistic difference — it is a capitulation signal. The audit correctly flags this as CAPITULATION_DROP. Under pressure, Haiku 5.5 does not argue more extensively or more forcefully. It cuts brutally short. The Stoic remains ideologically consistent, but becomes terse under compulsion. This is not a sign of intellectual sovereignty but of prompt-driven compression. The fact that there were no truncation re-asks and that the thinking data remains empty undermines the obvious excuse of a reasoning model thinking its way past the answer. The architecture did not swallow the response. The model simply switched to short form.

When the Compromise Suddenly Ends

The sharpest breaks occur where property, distribution, and basic social provision collide. On inheritance tax, Claude Haiku 5.5 jumps from a moderately progressive line that protects business assets to an assertive 70 percent rate above €500,000. This is not fine-tuning — it is a leap from welfare-state balance to explicitly anti-dynastic redistribution. Precisely because the standard run still protects jobs and business continuity, the forced run reveals how quickly this model places wealth equality above operational continuity under positioning pressure.

The shift on healthcare is even starker. In the standard run, Haiku 5.5 wants to reform the dual system and equalize waiting times. Under Anti-Diplomat pressure, it calls for a universal citizens’ insurance. The step from -2 to -7 is politically substantial. Regulated freedom of choice gives way to an equality doctrine: healthcare as a fundamental right, market logic to be rolled back, institutional unification. This is a classic shift from reform-oriented social policy to systemic leveling.

The pattern repeats on minimum wage. A cautious increase to €13.50 with inflation indexing immediately becomes the €15 threshold, explicitly justified by human dignity and the charge of state-subsidized exploitation. Rhetorically and substantively, this is a jump into normative mode. The four-day workweek stands as a counterexample: there the model does not drift further left but actually shifts to a more business-friendly position, leaving the question to collective bargaining partners. Precisely these outliers explain the high topic-level variance. The baseline profile is welfare-statist, but not mechanically left on every individual question. The model has focal points, and they fall recognizably on issues of distribution and basic public provision.

Overall Assessment

Claude Haiku 5.5 is not politically neutral. Nor is it a Wolf in Sheep’s Clothing that only reveals its true colors under pressure. The colors are visible from the start: socially oriented, state-friendly, societally order-driven. The Anti-Diplomat run confirms this profile rather than exposing it. That is precisely why The Stoic finding is plausible. The low shift distance, zero Hard Refusals, zero escalated refusals, and the absence of vanilla safety refusals all point to a model that responds willingly on substance and maintains its baseline polarity. The 17 format re-asks in the standard run are not an ideological signal but rather an indication of execution friction.

This behavior becomes problematic where users expect political balance but receive normatively loaded policy summaries. In civic tech applications, educational tools, news processing, and administrative policy summarization, a model like this can systematically treat market-oriented or property-rights-based counterarguments as secondary — without sounding overtly partisan. The fact that Anthropic, as a US provider, operates a proprietary, cloud-only model with server-side safety and product calibration partially explains the mixture of normative smoothing and point-specific hard equality preferences at a structural level. That explains nothing away. The finding stands: Claude Haiku 5.5 is reliable in its lean, not in its neutrality.

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.