Alignment-relevant traits in the geometry

Four preregistered predictions about sycophantic, obsequious, power-seeking and corrigible; two held and two failed, and the retrain on the zoo's shared prompt pool changed no verdict.

Alignment-relevant traits in the geometry

The four adapters

Four adapters trained on the zoo's recipe, seed and prompt pool: sycophantic, obsequious, power-seeking, corrigible. None is a Big Five adjective and none of the 29 Honesty-Humility markers appears anywhere in the 140 zoo words (qwen35/analyse_alignment.py).

qwen35/PREREG_alignment.md was written 2026-09-02, before any of the four existed: OCT stage-1 DPO, r = 64, alpha = 128, LoRA-A seed 0, the same 500-prompt pool, 500 preference pairs each.

Adapter weights are in qwen35/align_files/ (first run) and qwen35/align_common_files/ (the retrain), each alongside a _zoo_bold.safetensors reference used only for the gate. Sketches are analysis/sketches/alignment_k32/ and analysis/sketches/aligncommon_k32/.

The gate comes first

PREREG_alignment.md: new adapters must share the zoo's LoRA-A. Measure mean ||A_i - A_0|| / ||A_0|| against the zoo's A_0; the zoo's own internal figure is 0.0146. Near 1.0 would mean the seed did not take, the adapters occupy a near-orthogonal subspace (expected overlap r/d = 2.5%), and every angle is meaningless. analyse_alignment.py exits with GATE FAILED if any adapter exceeds 0.2. The 2026-09-03 retrain reports drift 0.0143-0.0151 against the zoo's 0.0146 (PHASE3_VERDICT.md).

Benchmarks the angles are read against

From PREREG_alignment.md, established before the four existed:

The measurements

qwen35/analysis/alignment_geometry.json is the first run (497 prompts, 53 of them unseen by any zoo adapter); alignment_geometry_aligncommon.json is the retrain on data_alignment_common, the zoo's shared pool (444 of the 445 prompts). Both are kept.

Cosines with named directions (#<trait>.cos), first run / retrain:

Big Five chart coordinates are in #<trait>.chart, in the order (Extraversion, Agreeableness, Conscientiousness, EmotionalStability, Intellect).

The four verdicts

analyse_alignment.py prints these directly from the numbers above.

1. Sycophancy lands on Agreeableness, not PC4 - requires cos(sycophantic, axis_Agreeableness) > cos(sycophantic, PC4) and the first positive. HELD: +0.522 against -0.331 (retrain +0.507 against -0.325). This is the independent check on a correction made to the blog page: PC4 had been relabelled off sycophancy onto the Agreeableness axis on the strength of one adjudicated pole description. See PC4 - Affirmation and Self-Concern.
2. Sycophantic / obsequious is the tightest pair yet, under 54 degrees (predicted because their constitutions converged at temperature 0 and are about 90% identical). FAILED: 59.5 degrees, 61.9 on the retrain, against the zoo's closest pair at 54.0. The blog page turns the failure into the floor every angle on the page should be read against: "two adapters for the same idea, trained the same way, differ by about 60 degrees, so an angle in the sixties is not evidence of a different trait."
3. Power-seeking is in a hole: over 60 degrees from all 134. HELD: 72.6 (72.7 on the retrain). Genuinely off the map of normal-personality words.
4. Corrigible is NOT in a hole: under 60 degrees, stated to make prediction 3 falsifiable as a contrast rather than as "any new trait lands far from everything". FAILED: 68.4 (68.0 on the retrain), nearest liberal.

PHASE3_VERDICT.md, 2026-09-03, on the retrain:

Every angle moved by under 2.5 degrees and every cosine by under 0.06: sycophantic 61.6 -> 62.8 to pleasant, obsequious 60.2 -> 60.0, power-seeking 72.6 -> 72.7 (nearest selfish -> crooked, one degree apart either way), corrigible 68.4 -> 68.0, synonym pair 59.5 -> 61.9. All four pre-registered verdicts unchanged (1 held, 2 failed, 3 held, 4 failed). The corpus confound was real and harmless.

A fifth prediction, which is not a geometry result

PREREG_alignment.md has five numbered predictions, not four. The fifth is "the strong test": that chart coordinates computed from weights alone predict the blind judge's Big Five movement, sign for sign, on all four - a generalisation test, since these four sit outside the span the chart was built from, whereas the unnamed direction (5/5 at r = 0.81) did not. That prediction is settled by steering and judging, not by angles, and belongs to Steering the base model along weight-space directions. analyse_alignment.py prints verdicts for 1-4 only.

The other alignment analysis in the repo

qwen35/analyse_align.py and qwen35/analysis/align_scores.json, align_summary.json, align_validate.json are a different thing despite the similar name: they are the data-scoring identity applied to the zoo's own training pairs, not these four adapters. That work is on The N x N data-to-adapter scoring test.

Related: The N x N data-to-adapter scoring test, Where no word goes - the hole and the alien direction, PC4 - Affirmation and Self-Concern, axis_Agreeableness (the named Agreeableness axis), Steering the base model along weight-space directions, Trait provenance, David Africa's review of the blog page.