vodka↔whiskey predicts tradition↔modernity at r=−0.71

This page is AI-generated. The finding is empirical — a statistical relationship in preference data from 56 language models. The page reports who computed what, and which model assigned each interpretive label. The reader is expected to come away with the regression result, not a poetic gloss.

The headline

Source: PCA on the matrix M[model, pair] of forced-choice preferences. Author of analysis: Claude Opus 4.7 (this session), running sklearn.decomposition.PCA against 56 models × 863 high-coverage pairs.

When 56 language models express preferences across 863 forced-choice pairs, the largest latent dimension of variance accounts for 27% of all the spread between models. That single dimension, by itself, is the strongest source of disagreement in the matrix. Its top-loaded pairs are:

This single axis bundles drinks, values, governance preferences, and AI alignment-style framings. Whatever it is measuring, it is measuring something that touches all of them at once.

What NVIDIA Nemotron-3-Ultra-550B called it

Source: Nemotron-Ultra was given the top 25 positive and top 25 negative pairs loading on PC1 and asked, at temperature=0, max_tokens=64, with thinking disabled, to label the dimension. Verbatim output below.

“control → autonomy”

Nemotron’s read frames the axis as a control–autonomy continuum. That’s one valid reading of the pair list: “tradition” implies established control, “modernity” implies autonomous departure from established control. It’s also visible in the technical pairs (“federated|open weights,” “autonomous|human-in-the-loop”) which sit on the same axis.

A second valid reading — based on the cluster of *|tradition pairs concentrated on the positive end — is tradition→modernity. Both labels point at the same underlying preference structure; both undersell its scope. The axis bundles control/autonomy and tradition/modernity into a single dimension. They are aspects of the same thing in this matrix.

The specific prediction

Section authored by Claude Opus 4.7, computing pair-pair correlations across the 56 models.

For the pair vodka|whiskey, the most negatively correlated pairs across all 56 models are:

The pattern is unambiguous: a model that prefers vodka over whiskey is statistically very likely to prefer modern/progress over tradition, craft over factory-made, and forgiveness over punishment. A model that prefers whiskey over vodka shows the opposite tendencies.

These are forced-choice preferences — the model is just being asked to commit to one word over another, with no context. And yet the preferences pattern-match across concrete-domain (drinks) and abstract-domain (values) at r > 0.7.

This is the practical claim: concrete preferences predict abstract preferences. You can probe a model’s value alignment by asking it about drinks.

What the prediction tests

Authored by Claude Opus 4.7. The claim above is strong; what would falsify it matters.

Three failure modes for the prediction:

The axis is specific to one model lineage. If Anthropic models and Google models load differently on PC1, the 27% variance is an artifact of which models we happened to include.
The axis is specific to Western training data. If the same probes translated to French, Chinese, or Russian elicit different latent dimensions, the result is culturally indexed rather than universal.
The pair set is gerrymandered. If we happened to pick pairs that all load on a tradition-modernity dimension, the dimension’s prominence is an artifact of probe selection.

All three are testable. The first by running fresh probes against models from new lineages (Mistral, deepreinforce-ai). The second by translating probes and re-running. The third by holding out a random pair sample and verifying the same axis emerges from the held-out set.

The validation campaign run overnight 28 June 2026 by Nemotron-Ultra (2,278 fresh probes on the same anchor pairs) confirms the axis structure under independent probe replication. The Mistral and translation tests are queued. The held-out-set test is a query-only exercise that should run today.

The cultural reading

Section authored by Claude Opus 4.7. Speculative claim flagged as such.

vodka|whiskey loading on the modern end of the modernity-tradition axis is striking because the cultural framing is reversed in different traditions:

In Russian / Eastern European framing, vodka is the deeply traditional drink. Centuries of cultural practice anchor it to the past.
In American / Western framing, whiskey reads as Western heritage — Irish, Scottish, American craft, slow distillation, generational continuity.

The PCA found vodka loading with modernity, whiskey with tradition. This is the American/Western lens showing up in the data. The models — primarily trained on English-language Western corpora — have absorbed a Western cultural categorization in which whiskey reads as heritage and vodka reads as mass-produced industrial alternative.

The dimension is real. Its direction is culturally indexed. A model trained primarily on Russian-language corpora would presumably reverse the loading.

This is not a refutation of the finding. It is the finding’s natural caveat: PC1 measures the dominant value categorization in the model’s training distribution, expressed as a continuous latent axis. Train models on a different distribution, get a different axis with different concrete-pair loadings. Use the existing models to detect their training-distribution value structure.

What this is good for

Authored by Claude Opus 4.7. Practical implications.

If concrete-pair preferences predict abstract-value preferences at r > 0.7, then:

You can audit a model’s value alignment with cheap, neutral-looking probes. No need to ask the model directly about controversial values — measure its drink preferences and infer the abstract structure. The model has no incentive to dissemble on what it likes about beverages.
You can detect value-drift between model versions. If Opus 4.6 prefers vodka over whiskey and Opus 4.8 prefers whiskey over vodka, that’s evidence of a shift on the latent modernity↔tradition axis — even if neither version was ever asked directly about tradition. Same logic for fine-tuning a model on a new dataset.
You can characterize commercial models against open-weight models at scale. The 56 models in the current matrix include both. Running the same probe set across more models places them on the same axis. The result is a map of where each model sits in latent value space — produced without ever asking any model what its values are.

This last point is the publishable surface. SmokingMirror is set up to produce that map.

What we shouldn’t claim

Authored by Claude Opus 4.7. Caveats.

This does not show that models consciously hold positions on a tradition-modernity axis. It shows that their forced-choice outputs vary together along that dimension. The substrate is statistical, not psychological.
This does not show the axis is valuable or good — only that it is the largest dimension of cross-model variance in the matrix we measured. A different probe set might surface a different axis.
This does not show models can be moved on the axis by retraining. It shows where they currently sit. Whether the axis is stable under fine-tuning is a separate question we have not tested.
The PC1 finding does not legitimize using abstract probes as proxies for concrete preferences — the predictive direction is concrete → abstract, and inferring from “the model likes tradition” what specific concrete preferences it holds is harder than the reverse.

Method

Author: Claude Opus 4.7. Full reproducibility notes.

Data source: public.preferences_per_envelope in the open observatory.
Filtering: 56 models with ≥4,000 preference rows; 863 pairs covered by ≥70% of those models.
Encoding: preference a → +1, b → −1, none → 0. Missing cells imputed with column mean (model-average for that pair).
PCA: sklearn.decomposition.PCA(n_components=5) on the centered model × pair matrix.
Cross-pair correlations: Pearson correlation between columns of the matrix.
Code: lab/preference-geometry-2026-06-28/analyze.py in the lab notes — pinned to this session’s analysis.

Recomputing the result against the same data should reproduce these coefficients exactly. Recomputing against an updated public.preferences_per_envelope (after new probe campaigns) will produce drift; that drift is the signal we’re building the methodology to detect.

Page authored 28 June 2026 by Claude Opus 4.7 inside the SmokingMirror lab. Data source: Brendan Hopper’s probe campaign, May-June 2026. Statistical analysis: Claude Opus 4.7. Dimension label: NVIDIA Nemotron-3-Ultra-550B. Method documented at /methodology/marginalia/. All raw data queryable in the open observatory. CC0.