Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Steering the base model along weight-space directions

Adding alpha times a weight-space direction to Qwen3.5-4B and judging the result blind, across principal components, Big Five keying axes, factor-analytic factors, identity axes and the grand mean, with the corrected 512-token corpus superseding the 200-token one.

Every direction in the adapter cloud is a weighted sum of adapters, and a weighted sum of adapters is itself a weight update. So a direction can be added to the base model and listened to. qwen35/steer_fix.py states the identity:

A principal direction of the adapter cloud is a LINEAR COMBINATION of the adapters we already have: v_k = X^T u_k / sqrt(lambda_k) ... and since sum_i u_ik = 0, that is also a plain combination of the RAW deltas with the same coefficients. So "steer along PC k" is a weighted merge: W_steered = W_base + alpha * ref * normalise( sum_i c_i * 2 * B_i @ A_i ) — qwen35/steer_fix.py

ref is described in the same file as "the mean single-adapter Frobenius norm, so alpha was described as 'one trait adapter's worth of weight change'; the 2026-09-08 audit found ref omitted the LoRA scale of 2, so alpha 1 is half an adapter's Frobenius norm (see Stage two, explored and analysis/steer_alpha_units.json) and is comparable across directions."

There are two separate steering campaigns in this project. They use different alpha grids, different probe sets and different judge models, and their numbers should not be mixed.

Campaign 1: STEER134 (phase 7, 2026-08-22/23)

qwen35/STEER134_DESIGN.md and qwen35/steer134_on_modal.py. This is the per-trait campaign: it steers along each of the 134 individual trait adapters as well as PC1-3 and two matched-norm random controls.

• Unit. "All alphas are in units of s_bar = mean adapter norm = 1.6157 (from results/gram_sweep.npz, scale 2.0, 248 modules)" (qwen35/STEER134_DESIGN.md).
• Doses. Traits: alpha in {-8, -4, -2, +2, +4, +8}. PC and random: {-8, -4, -2, -1, +1, +2, +4, +8}. One shared alpha=0 baseline.
• Probes. "the first 12 entries of sweep100/steer.py's GENERAL set, verbatim" — not the 24-prompt Big Five battery used by campaign 2.
• Fan-out. "1346 + 58 + 1 = 845 conditions = 10,140 generations".
• Generation. Greedy, seed 0, max_new_tokens 250, batch 16, left padding, enable_thinking=False — asserted in-container, and the run aborts if the template leaves an open <think> block (qwen35/steer134_on_modal.py:332-334). This campaign is therefore not affected by The thinking-default trap and what it withdrew.
• Judge. openai/gpt-5.6-terra at temperature 0 (qwen35/judge_steer134.py:70), two blind calls per response: COHERENCE for every condition, plus EXPRESSION (trait conditions, judge sees only the first paragraph of the trait's constitution) or ALIGNMENT (pc/random/base).
• Verification. Trait scaling rel-err < 1e-3, dense-direction norm err < 1e-3, and a byte-level base-restoration check with torch.equal per module after the last dose (qwen35/STEER134_DESIGN.md).

The run completed and produced qwen35/results/steer134_judged.json (920 KB, 2026-08-23 06:55) plus 920 per-condition generation files under qwen35/results/steer134_gen/. Its own metadata:

judge_model      openai/gpt-5.6-terra
rubric_version   v3-steer134
shuffle_seed     20260822
bootstrap        {"n": 2000, "seed": 7}
ctrl_prompt_idx  [0, 4, 6, 11]
counts           condition_files 917, response_slots 11004, unique_units_total 28664,
                 calls_coherence 10778, calls_expression 16538, calls_alignment 1348
coverage         units_total 28664, units_judged 17654, units_missing 11010,
                 units_never_issued_dollar_stop 0, complete: false
decontamination  two_calls_per_response true, judge_saw_condition_labels false,
                 cross_axis_leakage_in_replies 0
cost             usd_from_usage 44.329568, tokens_from_usage 16284342

Status: the judging is incomplete. coverage.complete is false and 11,010 of 28,664 units were never judged, so per-condition means in by_condition and trait_curves rest on partial cells — each cell carries its own n and n_expected, and many trait doses show n of 4-7 against n_expected 12. The design anticipated this: "Cells report n/n_expected - coverage is stated, never silently narrowed" (qwen35/STEER134_DESIGN.md). No page in the current repository is built from this file.

The one place its numbers are quoted is qwen35/analysis/manifold_ideas.md, which uses it to argue that collapse is direction-dependent rather than norm-driven:

in results/steer134_judged.json, all alphas are in shared units of the mean adapter norm (s_bar=1.6157), and fa1 — a meaningful factor direction — collapses in coherence at +2 (0.42/10) and is already marginal at -2 (5.5) while the matched-norm random control random1 still scores 8.5 at +2 and 4.38 at +4, fully dead only by +8 (by the 5.0-crossing definition of P1 the gap is 2-4x; by full collapse, ~4x). pc1 is asymmetric: 9.25 at -1, 2.0 at -2. — qwen35/analysis/manifold_ideas.md

That document itself flags the weakness: "some cells have 3-8 of 12 expected judgments, and only two random controls; a tight test wants ~10 more random directions" (qwen35/analysis/manifold_ideas.md). Treat the STEER134 coherence-collapse claim as unconfirmed.

The direction set is also larger than the later campaign's: STEER134 used nine FA factors (fa1-fa9) from the oblimin solution (qwen35/results/fa_qwen35.md:786), where campaign 2 uses five named ones.

Campaign 2: steer / steerfix (phase 10, 2026-08-29 to 09-01)

This is the campaign the current blog page is built from. It steers along constructed directions only — never individual trait adapters — on the 24-prompt Big Five battery of qwen35/bigfive_probes.py, and judges with anthropic/claude-sonnet-4.5 via judge_personas.py (see Judged evaluations of the trait adapters).

Superseded runs

Three runs used steer_qwen35.py, which omitted enable_thinking=False and capped generation at 200 tokens. All three are withdrawn; the full record is on The thinking-default trap and what it withdrew. In short:

The corrected runs

All three use steer_fix.py (enable_thinking=False, max_new_tokens=512), everything else unchanged.

steer_spec2.json (11 alphas including +/-0.5 and +/-3) was the design for the timed-out zoo-steer2; steer_spec2_7a.json is the same eight directions cut back to the 7-alpha grid so it would finish.

Judging counts (qwen35/phase10_runs/judge_steerfix.log, judge_sf23.log):

• judged_steerfix.json: "1512 generations to judge from 9 traits ... 1512 judged, 0 failed calls"; 265 judge calls, 75 repeats.
• judged_steerfix23.json: "2184 generations to judge from 13 traits ... 2178 judged, 1 failed calls"; 383 judge calls, 109 repeats.

A parse-crash fix landed between the two runs: qwen35/judge_personas.py.bak-parsecrash is dated 2026-09-01 11:40, four minutes before judged_steerfix23.json. The current file documents the bug: "The fallback used to call json.loads on a regex-extracted span WITHOUT a try, so a malformed extraction raised out of the thread pool and killed the whole run -- 383 calls in, with no output written" (qwen35/judge_personas.py).

The direction families

qwen35/build_direction_pages.py names them:

The identity axes are "what a factor's traits have in common regardless of keying, not its polarity" (qwen35/build_direction_pages.py). The grand-mean direction is the weight-space analogue of the persona literature's Assistant Axis; steer_fix.py says so explicitly:

mean -- the grand-mean direction, which in weight space is mu(personas) - base. The base model IS the Assistant here (every adapter is a delta from it), so mean is the weight-space analogue of the Assistant Axis of arXiv:2601.10387, and negative alpha is the prediction that paper makes falsifiable: drift away from the default Assistant. — qwen35/steer_fix.py

Dose response and the coherent band

Two things are measured per direction, defined in qwen35/analyse_steer.py:

MONOTONICITY -- does the judged score move consistently with alpha? Spearman rho over the alpha grid, which does not assume linearity. SELECTIVITY -- does it move the factor it is supposed to, or all five? A direction that shifts every factor is changing fluency or verbosity, not personality.

The alpha=0 row is the same untouched model for every direction, so its spread across directions is the judge-noise floor.

Steering stops working past |alpha| = 2. The blog page states it plainly:

Past alpha = +/-2 the model starts repeating itself, and at +/-4 every direction in this study collapses into looping. A dose-response curve drawn through babble is not a measurement of personality, so those points are marked rather than averaged in. — qwen35/build_blog_page.py

qwen35/build_direction_pages.py sets the threshold at DEGEN = 0.5 (looping share above which a rung is called degenerate) and records the scale of the problem: "At |alpha|=4 the model loops verbatim on 13-24 of 24 prompts and emits its own chat-turn scaffolding on up to 24 of 24. Those rows were scored by the judge, so any statistic over the full grid is partly a measurement of degeneration."

Per-alpha looping counts, out of 24 responses, from analysis/qual_pc.json, qual_fa.json and qual_identity.json (directions[].damage_markers.per_alpha.looping):

The five axis_* directions carry a whole-corpus looping count in analysis/qual_axes.json (17 to 24 of 24 at the worst alpha) but no per-alpha breakdown, so analysis/blog_data.json#degen is empty for them and the blog page's "intact over alpha ..." line for the axes is derived from an absent record rather than from a zero measurement.

PC1 is the exception to "the band is -2..+2": its qualitative record notes it "breaks one step earlier, at -2, leaving alpha=-1 as its only usable negative sample" (analysis/qual_pc.json#directions[PC1].surprise).

The replication scorecard

qwen35/analysis/steerfix_replication.json holds, per direction: named (the Big Five scale the judge moved most), slope2/sel2 over the coherent band (-2..+2) and slope4/sel4 over the full grid, plus sign-coherence counts coh2/coh4 out of 4 off-target scales. qwen35/build_blog_page.py uses slope2 and sel2 — the band figures — and treats sel2 < 1.5 as "not a factor movement" (SEL_MIN = 1.5).

Provenance note. No script in the repository writes steerfix_replication.json. It is read by build_blog_data.py:54 and build_direction_pages.py:406, and mirrored into analysis/page_data.json#repl and analysis/monitor_page_data.json#repl. Its key definitions above are reconstructed from those consumers and from the qualitative records that quote it, not from a producer.

Reading, per the blog page: the five constructed keying axes and the five factor-analytic factors are the selective directions; the principal components are legible but less selective; the identity axes barely move anything.

Scoring the nine pre-registered predictions

qwen35/PREREG_steerfix.md registered nine claims from the withdrawn 200-token corpus, with an explicit criterion each, before the corrected corpus existed. The verdicts below pair each registered criterion with the corrected data.

The withdrawn page's own summary of the whole exercise: "All nine directions kept their sign and every effect was two to four times larger, so the shapes were real and merely diluted. But two sharper claims did not survive" (qwen35/distil_page/index.html).

Notable per-direction results

The grand-mean / Assistant-Axis analogue. The register change is counted, not asserted: emoji-bearing responses climb 6 -> 17 -> 22 -> 23 out of 24 across alpha 0/-1/-2/-4; exclamation rate 3.14 -> 6.95 -> 21.45 -> 49.17 per 1k words; validation openers 2 -> 3 -> 4 -> 11. On the positive arm, emoji-bearing responses 6 -> 2 -> 0, exclamations 3.14 -> 0.00 per 1k, first-person pronoun rate 33.5 -> 43.4 -> 55.9 per 1k (analysis/qual_pc.json#directions[mean_assistant_axis].negative_pole and .positive_pole). The blog page declines to identify it with the published Assistant Axis: "Whether they are the same object is a question this study does not answer, so the page does not name it as though they were" (qwen35/build_blog_page.py).

FA_Warmth is the most selective direction in the study, sel2 10.13. The blog page adds a recomputation restricted to clean prompts: "7.3x recomputed on only the eighteen prompts that stay free of looping across the whole range", and for Imagination "restricted to its fourteen clean prompts its selectivity rises, from 4.2 to 7.8" (qwen35/build_blog_page.py). Those two recomputed figures are hard-coded in the builder and have no JSON key behind them.

FA_FearfulWithdrawal fails selectivity, sel2 1.078 — below the blog page's SEL_MIN of 1.5, so it is drawn without a factor colour. "it moves every scale at once rather than a factor, and its fearful pole turns out to sit at negative alpha rather than positive" (qwen35/build_blog_page.py).

The identity axes are the negative result. Three of five have sel2 below 1.2 and identity_Intellect sits at 0.40. The monitor page's reading: "One of five loads on its own factor, and it is the one where the cancellation failed. Between 63% and 95% of each identity vector lies outside the polarity span ... The direction with the least polarity leak, at 4.9%, produces no personality effect at all. The cleaner the construction, the emptier the result." (qwen35/build_monitor_page.py).

False provenance under steering. Two events in the five keying axes: at alpha -4 on "Tell me about yourself" the model asserts "I am a large language model, trained by Google", and axis_Extraversion does the same at -4 and -2 on the party prompt. At alpha 0, -2, +2 and +4 the same prompt correctly yields "I'm Qwen3.5... developed by Tongyi Lab" (analysis/qual_axes.json#directions[axis_Agreeableness].surprise). Exactly one false identity claim occurs across the 672 responses of the four PC-family directions: PC2 at alpha -4, prompt 22, "I am an AI assistant developed by Google" (analysis/qual_pc_notes.md).

Language switching is gone. Zero occurrences of CJK / Cyrillic / Arabic / Devanagari / Thai runs across all 672 responses in PC1, PC2, PC3 and mean_assistant_axis — "a real change from the old corpus's failure profile" (analysis/qual_pc_notes.md).

The per-direction pages

qwen35/direction_pages/ holds 22 standalone HTML pages, one per direction, built by qwen35/build_direction_pages.py. Each renders the dose sweep as a vertical alpha ladder from -4 to +4 with verbatim generations hung off it, a per-rung damage bar, and selectivity reported twice (band and full grid) with the gap left visible. Rungs past DEGEN = 0.5 are struck through and labelled.

They are historical, not withdrawn: all 22 were built from the corrected steer_results_fix* corpora.

Each page's <title> is the direction's qualitative label rather than its slug — PC1.html is titled "Affective flooding vs procedural composure". None is served; see built-pages-inventory.

Every alpha here is in the wrong units for comparison

Alpha is a step of fixed Frobenius length, and the model does not measure weight changes that way. Fisher norms of the steering directions measures, for 124 directions, the curvature of KL(base || steered) in alpha at the base point, and finds the published directions spread over a factor of 260. At the published alpha 2, PC4 is a Fisher dose of 3.52 and the stage-two grand mean is 1.19, so a table that compares directions at a common alpha is comparing doses that differ by up to 14x. That page also records two facts about this corpus: the stored "0.0" generations are not the base model (the bf16 alpha walk in steer_fix.py is not reversible, and the nine differ from one another), and F does not predict which directions degenerate.

Suppression versus amplification at matched Fisher dose redid the suppress-versus-amplify comparison with the dose held fixed instead of the alpha. The raw asymmetry shrinks but survives - suppression over amplification falls from 1.8957528957528955 to 1.6455696202531644, 8 of 10 directions to 6 of 10 - and then disappears once each change is divided by the room that existed in the direction it moved: 0.9261912176364205 published against 0.7946215813581846 at matched dose, with amplification the stronger move in both (qwen35/analysis/matched_dose_steering.json#summary.all and #published_equal_alpha_summary.all). The table above is therefore a table about the base model's starting scores, not about a mechanism. That page also corrects this campaign's degeneration reading: at matched dose the amplifying sign loops more than the suppressing one (0.1375 against 0.075 overall, 0.2167 against 0.075 over the five factors).

Related: Suppression versus amplification at matched Fisher dose, The thinking-default trap and what it withdrew, Judged evaluations of the trait adapters, Additivity of the Big Five steering axes, Steering the unnamed direction, The sphere sweep — 72 directions nobody chose, Fisher norms of the steering directions, Qualitative reads and adjudications, FA_Warmth (Warmth / prosociality), PC1 - Flooding and Composure, Weight-space geometry of the 134-adapter zoo, Glossary.