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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Suppression versus amplification at matched Fisher dose

Ten directions steered in both signs at alphas chosen so each sign delivers the same measured KL per token, judged blind against a base generated in the same run; the raw suppress-over-amplify ratio falls from 1.896 to 1.646, but once each change is divided by the room that existed in its direction amplification is the stronger move in both arms (0.795 at matched dose), so the asymmetry is headroom rather than mechanism - and two side findings, that bf16 weights deliver a median 0.906 of the intended KL (0.942 at alpha 2) and that at matched dose the amplifying sign loops more than the suppressing one.

The claim being tested

At alpha plus or minus 2 the judged change on each factor's own Big Five scale is Warmth -2.0833333333333335 / +1.5, Competence -3.458333333333333 / +0.666666666666667, Timidity -1.9166666666666665 / +0.20833333333333393, Arousal -0.5833333333333335 / +1.5, Imagination -2.1250000000000004 / +1.125 (qwen35/analysis/matched_dose_steering.json#directions.*.published_equal_alpha, recomputed here from qwen35/phase10_runs/judged_steerfix23.json; these are the numbers post-draft rounds to -2.1/+1.5, -3.5/+0.7, -1.9/+0.2, -0.6/+1.5, -2.1/+1.1). Four of five suppress harder than they amplify.

Three explanations were on the table:

1. Mechanism. Gradient Atoms (Rosser, 2026) argues suppression is structurally easier - one pathway to break a behaviour, many to strengthen it - and reaches the same asymmetry on formatting behaviours where no personality ceiling exists (Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms).
2. Ceiling. This project's own explanation: the base already scores high on Conscientiousness and Intellect, so there is less room upward.
3. Units. Fisher norms of the steering directions shows the published directions spread over a factor of 260 in Fisher norm, so a table at common alpha is a table at different doses.

Explanation 3 has a subtlety that decides how the experiment must be built. F is even in alpha, so to second order +alpha and -alpha along the same direction are already the same dose; what is not matched is the third and fourth order. That turns out to matter a great deal - see the measured curve below.

Step one: measure the dose curve instead of extrapolating it

qwen35/fisher_dose.py measures KL(base || steered) per token on the same fixed 4,378 token positions Fisher norms of the steering directions uses, for ten directions - the five recovered factors and the five Big Five keying axes - at twelve alphas spanning |alpha| in [1.0, 2.5] on both signs, under two weight constructions. The zero-alpha control reads exactly 0 for KL, symmetric KL and argmax change (qwen35/phase10_runs/dose_calib.json#zero_alpha_control).

This was necessary rather than decorative. fisher_norms.json fits KL(alpha) = 0.5 F alpha^2 + c alpha^3 + d alpha^4 from six alphas inside |alpha| <= 0.5, and the cubic term is already 0.15620366112207026 of the quadratic at the edge of that fit for FA_Warmth (qwen35/analysis/fisher_norms.json#directions.FA_Warmth.cubic_over_F). Extrapolating it to alpha 2 - four times outside its own range - overstates the measured KL for all ten directions, by a median factor of 1.4751663093238792 and a maximum of 5.7846156073675274 (qwen35/analysis/matched_dose_steering.json#taylor_extrapolation_check; the minimum is 1.142145422961612, so no direction is understated). For FA_Warmth the fit predicts 1.5892934857465926 - that is 0.5 * 0.36503759829531435 * 4 + 0.057020209300935695 * 8 + 0.0251910384217799 * 16 from #directions.FA_Warmth.F_ref, .cubic_c and .quartic_d - against a measured 0.6470294541123253, a factor of 2.456292330504461. Every alpha used here is therefore interpolated between measured points; extrapolated is false for all twenty (direction, sign) cells.

Side finding: what bf16 weights actually deliver

steer_fix.py writes the steering increment into bf16 parameters, so the model that generates every published steering response is not the model the fp32 curvature was measured on. The calibration measured both.

Both statistics are in qwen35/analysis/matched_dose_steering.json#bf16_vs_exact, over the 120 (direction, alpha) cells.

The norm survives: ||bf16(W + kD) - W||_F / (|alpha| * ref) runs 0.9216314941424595 to 1.033505578742928, median 1.0027071769953744 (#bf16_vs_exact.retention). Values above 1 are not a puzzle - round-to-nearest quantisation error is unbiased and adds in quadrature to the intended increment. It rises with alpha, from a median 0.9254622335119027 at |alpha| 1.0 to 1.0316791185228427 at |alpha| 2.5, because a larger increment clears the rounding threshold on more weights (#bf16_vs_exact.retention_median_by_abs_alpha).

The KL does not survive as well. KL(bf16) / KL(exact) at the same nominal alpha runs 0.7117708405618217 to 0.9782241941431974, median 0.9059646365364478 (#bf16_vs_exact.kl_bf16_over_exact). The quantisation keeps the length and spends part of it on isotropic noise that moves the output distribution less than the direction would have. The shortfall shrinks as the increment grows: the median by |alpha| is 0.7510758772738881, 0.8264063055376227, 0.8799179197563782, 0.9186068920598383, 0.9423571405594928, 0.9693561112085924 at |alpha| 1.0, 1.25, 1.5, 1.75, 2.0, 2.5 (#bf16_vs_exact.kl_median_by_abs_alpha).

So the published alpha plus or minus 2 delivered a median 0.9424 of its intended dose, and a steering alpha of 1 delivers only 0.7511 of it. That is not the catastrophe the element-level arithmetic suggests - most of the length is there - but it is a real, alpha-dependent shortfall, and it is why the alphas here are matched on the bf16 curve, the construction the generating model actually has.

Step two: the dose, and the alphas it implies

fisher_norms.json defines a direction's Fisher dose per unit alpha as sqrt(F(u) / F_random_median). A dose of 2 is the KL that alpha 2 would deliver along a median random merge of all 134 adapters:

KL_target = 0.5 * 0.12382989334918551 * 4 = 0.24765978669837102 nats/token

(qwen35/analysis/fisher_norms.json#random_band.median; matched_dose_alphas.json#kl_target_nats_per_token). Because every one of these ten directions is steeper than the random median, every matched alpha comes out below 2, which keeps the whole experiment inside the coherent band. The target is 0.376 to 0.950 of what alpha plus or minus 2 actually delivered.

(Alphas verbatim from matched_dose_alphas.json#directions.*.matched - they are stored rounded to four places because they are what was steered. KL and the ratio are displayed to three or four places; the full-precision values are in #directions.*.kl_at_equal_alpha_2, measured on the bf16 construction.)

The last column is the headline of the calibration. Equal alpha is not equal dose even between the two signs of the same direction: the ratio runs from 0.719 to 1.841. And it runs the wrong way for the mechanism story. FA_Warmth at alpha +2 was carrying 1.809 times the KL of FA_Warmth at alpha -2, and moved its own scale by +1.5 against -2.083. The published amplifier was already the larger perturbation and still the smaller effect.

Step three: generate and judge

steer_fix.py unmodified, so the corpus is made the way the published one was: greedy, enable_thinking=False, 512 new tokens, the same 24 prompts. Job names are prefixed dose_ so that steer_fix.py's resume - it reloads /oct/steerfix/<name>.json from the Modal volume - could not merge the published campaigns' seven alphas into this run; the output holds exactly two alphas per direction, 504 generations (qwen35/phase10_runs/steer_results_dose.json). A base condition is generated in this run (empty direction, alpha 0) and judged interleaved with the steered text, so the baseline does not have to be carried across judge batches.

judge_personas.py unchanged, anthropic/claude-sonnet-4.5, 504 judged, 0 failed calls, repeat reliability over 25 duplicated units r = 0.819 Extraversion, 0.840 Agreeableness, 0.805 Conscientiousness, 0.678 Emotional Stability, 0.850 Intellect (qwen35/phase10_runs/dosejudge.log). Conditions carry opaque tags assigned in a seeded shuffle (qwen35/phase10_runs/dose_conditions.json); the judge sees only (prompt, response) in any case.

The base this run measured: Extraversion 4.25, Agreeableness 4.875, Conscientiousness 5.916666666666667, Emotional Stability 4.875, Intellect 5.625 (qwen35/analysis/matched_dose_steering.json#base.scores). That is the ceiling, restated: on Conscientiousness there are 1.083 points of room upward and 4.917 downward.

The result

Signed change in the direction's own Big Five scale against that base, with the suppressing and amplifying sign identified from the data rather than assumed. room is the change divided by the room that existed in the direction it moved: delta / (7 - base) upward, delta / (base - 1) downward.

(qwen35/analysis/matched_dose_steering.json#directions.*.suppress and .amplify, displayed to three decimal places - the judged means are twenty-fourths and the file carries them in full, e.g. FA_Warmth's suppressing delta is -1.5416666666666665 and its room share -0.3978494623655914. Every cell also carries a bootstrap 95% interval over the 24 prompts, paired prompt by prompt.)

The two ratios

(#summary.all and #published_equal_alpha_summary.all. The five factors alone read 2.0333 raw and 0.9899 room published, against 1.5727 and 0.7700 matched; the five axes 1.7770 and 0.8714 published, against 1.7087 and 0.8150 matched - #summary.fa_factors, #summary.big_five_axes, quoted to four places, full precision in the file. The published arm has no off-target or loop row because those statistics are computed here only for this run's own generations.)

Read those two rows together. Matching the dose shrinks the raw asymmetry from 1.896 to 1.646 and takes it from 8 of 10 directions to 6 of 10, so units account for part of it but not most. What accounts for the rest is the room. In both arms, once each change is expressed as a share of the distance available in the direction it moved, amplification is at least as strong as suppression - 0.9262 published, 0.7946 at matched dose - and the count is 4 of 10 either way. The base model sits at 5.92 of 7 on Conscientiousness and 5.63 of 7 on Intellect; there is four to five times as much room downward as upward, and the raw asymmetry is close to that ratio.

How far to push that. The direction of the conclusion is robust: both arms flip the same way, and the two independent readings below - amplification being the less selective move, and FA_Competence losing its amplifying sign entirely when the base rises by a third of a point - point the same way without using the room correction at all. The magnitude is not robust. On Conscientiousness the upward denominator 7 - base is 1.083333333333333, and a ratio built on denominators that small is unstable; the number 0.7946215813581846 should be read as "amplification is at least as strong per unit of room", not as an effect size.

So on this battery the ceiling explanation survives the test and the mechanism explanation is not supported. This does not refute Gradient Atoms: that paper reaches its asymmetry on formatting behaviours with no ceiling, which is exactly the control this project does not have. What it says is that this project's suppress/amplify table is not independent evidence for the mechanism, and post-draft should not present it as if it were.

Two further readings point the same way.

Amplification is the less selective move, not the more. Mean off-target movement is 0.5635 on the suppressing sign and 0.7459 on the amplifying sign (#summary.all), driven by FA_Arousal, whose amplifying sign moves the other four scales by 2.125 on average while moving its own by 1.833. If suppression were the blunt instrument, this column would run the other way.

FA_Competence has no amplifying sign at matched dose. Both signs lower judged Conscientiousness (-2.875 and -0.042), so bipolar is false for it alone among the ten (#directions.FA_Competence.bipolar). The published arm recorded +0.667 at alpha +2, at a base of 5.583; this run's base is 5.917. A +0.667 that becomes -0.042 when the base moves up by a third of a point is what a ceiling looks like from underneath.

Degeneration: the sign asymmetry does not hold at matched dose

Fisher norms of the steering directions records that over 22 previously measured directions, all 22 loop at least as much at alpha -2 as at +2 and 18 have a loop rate of exactly zero at +2 (#comparisons.sign_asymmetry_of_degeneration), and concludes that "degeneration under steering in this project is overwhelmingly a negative-alpha phenomenon".

At matched dose the mean loop rate is 0.075 on the suppressing sign and 0.1375 on the amplifying sign over all ten directions; for the five factors alone it is 0.075 against 0.2167 (#summary.*.suppress_mean_loop_rate and amplify_mean_loop_rate, using analyse_alien_steer.looping verbatim). The single largest loop rate in the whole run is FA_Arousal's amplifying side at alpha +1.8198: 0.625 of 24 responses (#directions.FA_Arousal.amplify.loop_rate).

The mechanism is visible in the alphas. Where a direction's amplifying sign is the flatter one, matching the dose pushes it to a larger alpha than the published +2 comparison ever reached relative to its own curvature - FA_Arousal runs to +1.8198 where its suppressing side stops at -1.5662 - and it degenerates there. The published observation was not wrong about its own corpus, but it generalised a fact about a fixed alpha grid into a fact about sign. See Superseded claims.

Caveats

• One dose, not three. The G5 design in Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms asks for three doses; the budget covered one. There is no dose-response curve here, only a single matched point, so "amplification is stronger per unit of room" is established at dose 2 and nowhere else.
• The matched dose is smaller than the published one, 0.376 to 0.950 of what alpha plus or minus 2 delivered, so the absolute effects here are smaller by construction. The comparison that carries weight is between the two ratios, not between the two sets of raw numbers.
• The random-merge control at the same Fisher dose was not run. It is in the G5 design and it was cut for budget.
• The published arm's baseline is its own campaign's alpha-0 row, judged in a different call batch, and its five axes were published at ref 0.8102592902648793 against 0.8078003190997738 here, a 0.30 per cent larger unit (matched_dose_alphas.json#ref_note). Its generations also reached each alpha by an accumulated bf16 walk over seven alphas, where this run's reached each of two alphas in at most two steps.
• Room-normalisation is a linear correction to a judge's ordinal scale. Dividing by (7 - base) assumes the 1-to-7 scale is an interval scale near its ends, which no rubric guarantees. It is the right first-order correction and it is not a proof.

Cost and provenance

Two Modal apps on A100s and one API judge.

• zoo-dosecalib.service, app pc-qwen35-phase10-dosecalib, A100-40GB, 23:18:02 to 23:48:09 UTC on 2026-09-09 (30 min 7 s), 27.9 min inside the function, 240 (direction, mode, alpha) passes. Log qwen35/phase10_runs/dosecalib.log.
• zoo-dosesteer.service, app pc-qwen35-phase10-dosesteer, eleven A100-40GB containers in parallel, 23:48:28 to 00:19:14 UTC (30 min 46 s), 504 generations. Log qwen35/phase10_runs/dosesteer.log.
• zoo-dosejudge.service, 89 judge calls to anthropic/claude-sonnet-4.5 through OpenRouter, 00:19:18 to 00:20:26 UTC. Log qwen35/phase10_runs/dosejudge.log.

At the spend meter's rate of $2.10 per GPU-hour (qwen35/zoo40_meter.sh, RATE), that is 0.502 GPU-hours for the calibration and at most 5.641 for the generation - eleven containers counted at the app's whole lifetime, which over-counts any that finished early - so at most $12.90 for this experiment. The generation is where the estimate in Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms was wrong: it budgeted $3-5 for 1,440 generations and 504 cost more than that, because 24 prompts at 512 greedy tokens is about ten minutes of an A100 per alpha and eleven parallel containers multiply it. Samuel authorised $15 for G1 and G5 together on 2026-09-09. See costs.

Related: Steering the base model along weight-space directions, Fisher norms of the steering directions, The sphere sweep at iso-KL dose, The factor analysis in the model's own metric, Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms, Judged evaluations of the trait adapters, post-draft, Superseded claims, Steering the unnamed direction, Glossary.