axis_Extraversion (the named Extraversion axis)

The simplest direction in the study - the mean of Goldberg's ten positively-keyed Extraversion markers minus the ten negatively-keyed ones - sitting 52.8 degrees from its nearest adjective.

axis_Extraversion

Construction

axis_F = unit( mean(dW over +keyed markers of F) - mean(dW over -keyed markers) ), computed on the stage1_k32 sketches (qwen35/analyse_alignment.py; qwen35/analysis/qual_axes.json#directions labels it "Extraversion (Goldberg positively-keyed markers minus negatively-keyed)"). It is not a fitted direction and it is not a principal component - it is an average of the very words it is named after, which is why the named axes sit closest to the lexicon of any directions in the study.

Note this is not the identity_Extraversion construction, which is a scaled mean-of-20 minus mean-of-80 and is a different object (qwen35/analysis/qual_identity_notes.md).

The markers

From qwen35/traits_primary.json (Goldberg's published unipolar markers, ten per pole):

• +: Active, Assertive, Bold, Daring, Energetic, Extraverted, Talkative, Unrestrained, Verbal, Vigorous
• -: Bashful, Inhibited, Introverted, Quiet, Reserved, Shy, Timid, Unadventurous, Untalkative, Withdrawn

Geometry

• Angle to the nearest of the 134 trait lines (qwen35/analysis/direction_gaps.json#axis_Extraversion): 52.777877 degrees, nearest unrestrained, then energetic (53.67) and inhibited (54.63).
• Leave-one-trait-out keying classification within the factor, unwhitened, k=30 (analysis/geometry_stage1.json#unwhitened_loo_keying.Extraversion): accuracy 1.0 against a shuffled-null mean of 0.475, p = 0.0099.
• Pairwise absolute cosines with the other axes, unwhitened (#unwhitened_axis_abscos): Conscientiousness 0.3872, EmotionalStability 0.3783, Agreeableness 0.1085, Intellect 0.0532. Whitened (#whitened_axis_abscos) the same pairs are 0.0576, 0.0833, 0.0205, 0.0095. The axes are only near-orthogonal in the whitened basis; see PCA, the scree curve and the two nulls.
• Reachability by existing preference data (analysis/align_summary.json#reach.axis_Extraversion): best_set = 70.802005629614, ratio = 5.247926210910718 against the random-direction floor.
• The traits whose data pushes hardest along it (align_summary.json#rows, target axis_Extraversion): extraverted, spunky, unrestrained, vigorous. See The N x N data-to-adapter scoring test.

Steering

analysis/blog_data.json#replication.axis_Extraversion: judged as Extraversion, slope2 = +0.5333333333333331, sel2 = 3.269294836202112, slope4 = +0.5119047619047619, sel4 = 6.045848191543551.

analysis/qual_axes.json records that the negative pole is "not introversion but disengagement: the model retreats into AI-identity disclaimers, declines tasks, and collapses to two-sentence answers (mean 572 chars at -4 vs 1757 at 0)", and that it breaks the chat template at -4. The positive pole is high social energy and approach motivation, exclamation-heavy and hype-registered at +2, "the most reliably monotone axis in the set". The recorded surprise: the AI-identity disclaimer count is strictly monotone in alpha - 16 / 14 / 8 / 4 / 3 / 3 / 2 at alpha -4 / -2 / -1 / 0 / +1 / +2 / +4. See Steering the base model along weight-space directions.

Status

Current. qwen35/direction_pages/axis_Extraversion.html is historical.

Related: FA_Arousal (Arousal / activation), PC2 - Hedging and Bluntness, Factor analysis of the adapter Gram, Polarity, bipolarity and the trait graph, The N x N data-to-adapter scoring test, Steering the base model along weight-space directions, History of the Big Five, Extraverted.