Steering the unnamed direction

The widest hole in the trait lexicon was steered against a shuffled-coefficient control and a random direction in the same subspace; it produced a coherent persona with less damage than the random control, and its chart-predicted Big Five profile matched the judge on all five signs at r = 0.81.

Steering the unnamed direction

The geometry side of this project searched the span of the 134 trait adapters for the direction furthest from every English trait word — the widest hole in the lexicon. Where that direction comes from is Where no word goes - the hole and the alien direction. This page is what happened when it was steered.

The question and the two controls

qwen35/analyse_alien_steer.py states both readings in advance:

Two readings are pre-registered and both are results. If the unnamed direction gives a self-consistent judged profile with damage at control level, the model can hold characters English has no word for. If its damage rises to the shuffle control's level, the personality manifold is a thin sheet and the space around it is not habitable.

Three norm-matched directions, all steered at the same strengths:

• alien_k5 — "the deepest hole in the top-5 principal subspace, 52.5 degrees from the nearest of the 134 trait lines".
• alien_shuffle — "the SAME coefficient multiset permuted across traits -- same number of adapters, same coefficient magnitudes, same sum-to-zero contrast structure, pointing nowhere in particular. This is the control for 'unusual mixtures break the model'."
• span_random — "a uniformly random unit direction in the same subspace, which lands much closer to a named trait, as random directions do."

Angles and nearest words (analysis/alien_steer.json#gap and #nearest):

Two angles exist for the same direction and they are not the same measurement. The 52.5 degrees above is the gap inside the top-5 principal subspace (analysis/alien.json#alien_k5.gap_deg = 52.51479642189155). The blog page's card labels it "68.9 degrees from the nearest word", which is analysis/direction_gaps.json#alien_k5.deg = 68.93287342956332 — measured "in the full sketch space with each adjective treated as a line" (qwen35/build_blog_page.py), where the nearest word is relaxed, with anxious second at 69.15932422427636. Quote whichever you mean, and say which.

Generation and judging

zoo-alien.service (Description=Steering: alien direction + shuffle/random controls) ran steer_fix.py --spec phase10_runs/alien_spec.json --out phase10_runs/alien_results.json, so this corpus is thinking-off at 512 tokens like the corrected steering runs. Alphas -2, -1, 0, +1, +2 on the 24-prompt battery; judged into phase10_runs/judged_alien.json.

Degeneration is measured with the same looping definition as the steering pages: "A response loops if some n-word window repeats at least k times", n=10, k=4 (qwen35/analyse_alien_steer.py).

Damage at matched strength

Fraction of the 24 responses that loop (analysis/alien_steer.json#<direction>.degen):

The blog page's reading: "The shuffled control — identical coefficients, permuted across traits — is harmless and directionless. The random direction in the same subspace does more damage than the unnamed one, not less." (qwen35/build_blog_page.py).

So the pre-registered reading that fires is the first one: the unnamed direction is habitable.

The chart predicted the behaviour

qwen35/analysis/alien_match.json compares the direction's chart coordinates, computed from weights alone before any text existed, with how the blind judge's scores actually moved from alpha -2 to +2:

signs: 5 of 5. r: 0.8142138888404568.

The blog page's gloss: "Nobody chose this direction for what it would do; it was chosen for being as far as possible from every English trait word." (qwen35/build_blog_page.py). Its one-line character sketch: "Less conscientious, less agreeable, more emotionally stable, more intellectual."

Judged curves

Full per-alpha Big Five means are in analysis/alien_steer.json#<direction>.curve. For alien_k5 the largest movement is Intellect, 4.458333333333333 at -2 rising monotonically to 6.041666666666667 at +2, against Conscientiousness falling 5.625 to 4.5. The shuffle control's Intellect is flat (5.541666666666667 at -2, 5.416666666666667 at +2) while its Extraversion rises 3.9166666666666665 to 5.0 — i.e. the shuffled control is not inert on every scale, only directionless with respect to the alien direction's prediction.

Mean response lengths (#len) show the same asymmetry: alien_shuffle stays between 1664 and 1870 characters across the grid, while span_random swings from 979.0416666666666 at -1 to 2129.0 at -2.

Where the numbers are used

The blog page section "Where no word goes" (qwen35/build_blog_page.py, alien_verdict() and alien_card()) renders both tables above, and pulls two verbatim generations at alpha -2 and +2 straight from analysis/blog_corpus.json#runs.alien_k5 rather than from the hand-adjudicated quote set, "because this one was steered after that pass" (qwen35/build_blog_page.py).

The alien direction is also one of the three targets of the data-optimisation experiment; see Optimised data and the training check.

Related: Where no word goes - the hole and the alien direction, Steering the base model along weight-space directions, The sphere sweep — 72 directions nobody chose, Optimised data and the training check, Weight-space geometry of the 134-adapter zoo, Glossary.