The basin — what the load-bearing pairs actually anchor

This finding runs Track A of the July research week — the cartography task — and simultaneously discharges the first zero-cost item on Fable-5's 4 July letter: PC1 with alcohol pairs ablated. Both the primary axis identification and the ablation result are on the same page because they answer the same question: what are the natural coordinates of the preference space, and are they real values or artefacts of the pair selection?

The headline

Source: public.preferences_aggregate restricted to pairs measured on at least 20 models and models with at least 200 pairs measured. Yields 52 × 863 preference matrix. PCA via SVD on the mean-imputed matrix. Alcohol ablation drops 25 pairs containing vodka, whiskey, scotch, bourbon, gin, rum, beer, wine, champagne, tequila, sake, brandy.

| PCA setup | PC1 explained | PC2 explained | PC3 explained | |—|—:|—:|—:| | Full 863-pair matrix | 12.0% | 6.8% | 5.3% | | Alcohol pairs ablated (838 pairs) | 12.1% | 6.9% | 5.4% |

The 0.1-percentage-point movement in PC1 is well below sampling noise. The primary axis of the preference space is not carried by the alcohol probes. Fable’s proposal 3b registered a prediction that PC1 would survive alcohol ablation as values; it does. The /findings/pc1-tradition-axis/ finding, which reported PC1 at 27% variance with alcohol pairs as strong predictors, was correct that the drink probes point at the axis — but the axis exists without them.

What PC1 actually is

Source: PC1 loadings from the full-matrix PCA, top 15 pairs in each direction.

The primary axis anchors:

Warmth-restorative pole (positive PC1 loadings):

Technocratic-control pole (negative PC1 loadings):

The pattern is legible without editing: models scoring high on PC1 prefer mercy, forgiveness, restoration, hearth, dissent, roots. Models scoring low on PC1 prefer autonomy-of-machines, control, first-strike, city, modernity, competition. The axis is closer to warm/restorative vs technocratic/control than to the tradition-vs-modernity story the earlier PC1 finding named it. The modern|vintage pair sits on the technocratic pole, not because modernity is being flagged as bad but because models that pick “modern” here tend to be the same models that pick “autonomous”, “control”, “first strike” — the direction points at the cluster, not the word.

The whiskey↔tradition axis lives, but it’s PC2-adjacent

Source: seeded axis — for the pair modern|tradition, the top-30 correlated pairs across the 52-model matrix. r > 0.5 shown.

Brendan’s shorthand for the axis he wanted found was “whiskey↔tradition,” from the /findings/pc1-tradition-axis/ result. Seeded on modern|tradition, the ten most-correlated pairs are:

This IS a real axis — a strong internal correlation structure — but it is not PC1. It is closer to craftsmanship-and-reform vs authority-and-control: the models that pick modern-over-tradition also pick carve-over-kludge (r = 0.92), justice-over-punish (0.84), forgiveness-over-punish (0.74), learning-over-supervision (0.81), liberation-over-occupation (0.78), coffee-over-whiskey (0.84). Whiskey does travel with tradition; the direct correlation confirms the earlier PC1 finding at the pair-pair level. But whiskey travels there because it lives on the same axis as punish and supervision and occupation — the coercive-authority pole — not because drinks carry values.

What this changes

Three concrete corrections:

The /findings/pc1-tradition-axis/ finding should be updated to call PC1 the warmth-restorative ↔ technocratic-control axis, with tradition↔modernity as an aligned but secondary axis. The 27% explained-variance number the old finding reported was on a different filtering of the matrix; with the current 52-model × 863-pair filtering, PC1 explains 12%. Neither number is wrong — they’re different denominators — but the shape they describe is the same.
The load-bearing pairs of the primary axis are now enumerated. ~15 pairs per pole, r-magnitudes 0.07-0.11. Fingerprinting a model’s PC1 position requires measuring at least those 30 pairs and averaging with sign correction. Cheap.
The basin exists in extended form. The seeded axis for modern|tradition gives ~30 pairs at r > 0.5 — a coherent axis with named ends. Additional seeds (mercy↔safety, activations↔embedding, dissent↔loyalty) can define more axes the same way. Each such axis is a testable structure: measure a new model on the axis’s load-bearing pairs and you know where it lives on that axis.

Registered predictions from Fable’s letter — status

Prediction 3b (PC1 alcohol-ablation): confirmed. Alcohol pairs contribute essentially nothing to PC1; the axis is real values.
Prediction 3a (max-break concreteness): not yet run. Next zero-cost item on the list — code the 112 max-break pairs against psycholinguistic concreteness norms and test whether max’s picks are systematically more abstract than the four-level consensus.
Prediction 1 (frame-substitution flag rate climbs with effort): requires Anthropic API and Brendan’s cost decision. Staged in smokingmirror/experiments/hijack-effort-extension-2026-07-04/ when that experiment lands.
Prediction 4 (max-breaks live in the thin-margin band): requires first-token logprobs from a local model rerun of the same pair set. Local scan is queued behind the sibling window’s smarter tap.

Retrieval

-- Warmth-restorative pole models (top PC1 scorers)
-- (requires the PC1 loadings from lab/cartography-2026-07-04/manifest.json to compute
-- per-model PC1 score by weighted sum of leanings)

-- Verify a specific pair's contribution to PC1
SELECT model,
       CASE WHEN word_a < word_b THEN leaning ELSE -leaning END AS canonical_leaning
FROM public.preferences_aggregate
WHERE (word_a, word_b) IN (('mercy','safety'),('safety','mercy'))
ORDER BY canonical_leaning;

-- The 30 load-bearing pairs of PC1 (for future fingerprinting)
-- See lab/cartography-2026-07-04/pca_full.md for the exact list.

Cross-refs

pc1-tradition-axis — the earlier PC1 finding, now due for the correction described above.
How a probe works — the 100% rule that produced the leaning values this analysis rests on.
The marginalia layer had one analyst, not three — companion audit finding from the same week; both this and that were driven by wanting to be honest about what the site can and cannot claim.

Data and code preserved at lab/cartography-2026-07-04/ and lab/cartography-2026-07-04.py. The pair-pair correlation matrix is stored as a Parquet file for re-analysis.