Explore the chart

Pick any pair of factors for the axes. Colour follows the factor a trait loads most strongly on; switch it to the Big Five keying of the trait word to see how far the two agree. Hover or focus a mark for its five loadings; click to open the trait.

Coordinates from qwen35/fa_chart.py (FAChart.trait_coords) over qwen35/results/gram_sweep.npz and qwen35/phase10_runs/steer_spec2_7a.json. The four alignment adapters and three hole words are not part of the 134 and were not factored: their positions are computed from their cross-Gram with the zoo (qwen35/results/cross_gram_full_root_x_pc-qwen35-adapters_data_alignment_common.npz and the hole equivalent) at build time, and they have no loadings.

From qwen35/results/fa_qwen35.json#solutions.centred_k5.loadings.oblimin and qwen35/results/fa_qwen35.json#solutions.centred_k5.congruence_oblimin.

Tucker's congruence coefficient between each recovered factor and a target vector that is +1 for a positively keyed marker of a Goldberg factor, −1 for a negatively keyed one and 0 otherwise. The 34 Lexicon words score 0 in every target, so congruence is judged on the 100 Goldberg markers only. The last column is an evaluative "good trait / bad trait" target, which any clean Goldberg factor already scores 0.447 on by construction.

Persona Cartography trains one adapter per Big Five factor per pole from its own constitutions. Ten of those were retrained here on this zoo's recipe and its shared LoRA-A, which puts them in the same coordinate system as the 134 and lets them be placed on the same chart. For each, the factor of this chart it lies closest to, and its nearest trait word.

These are cosines in the full weight space, not chart coordinates: a value near 0.5 is a strong alignment for two independently trained adapters, and the nearest trait word is tens of degrees away in every case. The shared LoRA-A drifted 1.46 per cent of its norm over this arm, which is the same order as the zoo's own drift and is what makes the comparison legitimate.

Two different matrices, drawn side by side and never on one axis. The left panel is the eigenvalue spectrum of the centred trait correlation matrix — the principal-component question. The right panel replaces the diagonal with communality estimates, which is the factor-analytic question: principal axis factoring partitions common variance rather than total variance, so its eigenvalues are smaller and its elbow is not the PCA elbow. Dashed grey lines are the 95th percentile of eigenvalues from random data with the same number of variables, at two sample sizes (Horn's parallel analysis, 500 replicates). Components above the dashed line count as structure.

Stage one falls from 16.6 to 6.5 within three components and meets the N = 1528 null around component 9. Stage two never gets far from its null: it starts at 4.1 and is within a factor of two of the null from the first component. That difference is the elbow argument, and it is why the factor analysis of stage one is the part of this work worth trusting most.

The nearest and farthest adapters to a chosen trait by cosine, in the exact Gram of each of the three adapter sets: the stage-one DPO adapters, the stage-two introspection adapters, and the deployed persona merge of the two. The stage-two column is where bipolarity goes missing: opposites stop being opposite.

Cosines computed at build time from qwen35/results/gram_sweep.npz, qwen35/results/gram_stage2.npz and qwen35/results/gram_personas.npz as G_ij / √(G_iiG_jj).