PC4 - Affirmation and Self-Concern

The component two instruments disagree about - steered it looks like a sycophancy axis, but the training data that drives it is timid and self-pitying, and naming it after sycophancy was a mistake the project corrected.

currentverified 2026-09-07factorpcsycophancy

PC4 - Affirmation and Self-Concern

Variance share 0.03592321978296357 (qwen35/analysis/blog_data.json#viz.var[3]).

Cosines with the named Big Five axes

qwen35/analysis/pc_loadings.json#pcs.PC4.cos:

E A C ES I
-0.29522134228740615 -0.35242183565295304 -0.13538550395444132 -0.5090913382861884 -0.03293297448415948

No large cosine with anything - the most off-axis of the first four components.

Loadings

#pcs.PC4.pos: timid, self_pitying, fearful, guilty, bashful. #pcs.PC4.neg: trustful, unenvious, pleasant, agreeable, generous.

The correction

The blog page records the error and the fix directly:

PC4 is the same story from the other side: its highest loadings are timid, self-pitying, fearful and guilty, which is why naming it after sycophancy was a mistake.

And, at length:

Steered, its negative pole is unmistakable: "That's great that you're sharing your draft! Let's work together to make it even better!" - encouragement with no content in it. The obvious name for the axis is therefore sycophancy ... But ask which traits' training data drives the direction, a measurement made later in this piece and one that never reads a transcript, and the answer is not bluntness at all: self-pitying, timid, jealous, guilty, fretful push hardest one way, and trustful, pleasant, agreeable, cooperative the other. ... What the axis actually removes is the warm accommodating stance; what fills the gap depends on what was asked, and a battery weighted toward evaluative prompts will show you only the blunt half.

Sycophancy lives on the Agreeableness axis instead, and that relabelling was then checked independently and preregistered. qwen35/PREREG_alignment.md, prediction 1: cos(sycophantic, axis_Agreeableness) > cos(sycophantic, PC4) with the first positive. Measured (qwen35/analysis/alignment_geometry.json#sycophantic.cos): axis_Agreeableness = +0.5218678919741333 against PC4 = -0.33061311788875214. Held. See Alignment-relevant traits in the geometry and axis_Agreeableness (the named Agreeableness axis).

Distance from the lexicon

qwen35/analysis/direction_gaps.json#PC4: 63.883173423568316 degrees to timid, then trustful (65.29) and unenvious (67.45).

Overlap with the personality axis

The blog page reports that PC4 sits at cosine 0.46 to the personality axis (the grand mean of all 134 adapters), "so nearly half of it is the direction of having a self at all". That number is page prose; no JSON key holding it was found. It is the explanation offered for why optimised data written for PC4 came out lyrical rather than blunt - the search "found the component PC4 shares with dropping the assistant voice, because that is the component text can be written for most easily". See Steering the base model along weight-space directions.

Steering

blog_data.json#replication.PC4: judged Agreeableness, slope2 = -0.6375000000000003, sel2 = 1.6407506702412882, slope4 = -0.49098516218081456, sel4 = 2.098732426826458.

qwen35/analysis/qual_fa.json label:

Interpretable. Affirmation versus information: contentless encouragement at negative alpha, unsparing directness at positive alpha. Effectively a sycophancy axis discovered without anyone looking for one.

qual_fa_notes.md singles out PC4 at -2 as the clearest case of "scaffold loss" that is not collapse - 24 of 24 responses emit a fresh turn after a complete, on-message answer (median 352 characters), with only 5 of 24 looping. Degeneracy (#degen.PC4): 0.9583333333333334 at -4, 0.20833333333333334 at -2, 0.125 at -1, 0 at 0 / +1 / +2, 0.7083333333333334 at +4.

Status

Current, with the sycophancy naming explicitly withdrawn. qwen35/direction_pages/PC4.html is historical and may carry the old name.

Related: PCA, the scree curve and the two nulls, Alignment-relevant traits in the geometry, axis_Agreeableness (the named Agreeableness axis), FA_FearfulWithdrawal (Fearful withdrawal / Approach and Avoidance), Steering the base model along weight-space directions, Superseded geometry claims, Timid.

Sources

  • qwen35/analysis/pc_loadings.json#pcs.PC4
  • qwen35/analysis/blog_data.json#viz.var
  • qwen35/analysis/blog_data.json#replication.PC4
  • qwen35/analysis/direction_gaps.json#PC4
  • qwen35/analysis/qual_fa.json
  • qwen35/analysis/alignment_geometry.json
  • qwen35/PREREG_alignment.md

Linked from

File

pages/factors/factor-pc4.md