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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Adapter effect and LoRA-A drift

The shared LoRA-A moves 1.5% of its norm across the whole stage-one zoo and about 10% in stage two, which is what makes the stage-one geometry comparable and weakens the stage-two case.

Why A drift is the load-bearing number

A LoRA writes dW = s * B @ A with B initialised at zero. Because B starts at zero, A receives almost no gradient, so it stays near its random initialisation. Every adapter in the stage-one zoo shares one seed, so if A really does stay put, all 134 write into the same random 64-dimensional slice of a 2560-dimensional input space and are directly comparable. If A moved a lot, they would not be.

That is the assumption on which every angle, cosine and principal component in Weight-space geometry of the 134-adapter zoo rests, and it is measured rather than assumed.

Stage one: 1.5%

Measured value: mean ||A_i - A_0|| / ||A_0|| = 0.014605041334818797, stored as qwen35/analysis/align_summary.json#a_drift and computed inside qwen35/align_score.py (_build_targets accumulates (A - a0).norm() / a0.norm() over modules).

Downstream uses of the same figure:

• The blog page rounds it: "across the whole zoo, A moves just 1.5% of its norm during training."
• qwen35/analyse_alignment.py and qwen35/analyse_hole.py both use "the zoo's own internal figure 0.0146" as the gate for new adapters, failing loudly if a new adapter's drift reaches 0.2 (the failure it is designed to catch is a seed that did not take, which "would read as a finding"). The 2026-09-03 alignment retrains came in at 0.0143-0.0151 (PHASE3_VERDICT.md); the three hole-word adapters gated at 0.013-0.015 (.garden/journal/2026-09-04.md).
• qwen35/probe_invariant.py gives the complementary measurement across seeds: "measured row-space overlap 0.0249, which is 64/2560, exactly what two random rank-64 subspaces of a 2560-dimensional space give."

A contradicting figure exists. qwen35/build_monitor_page.py line 112 says A "drifts 4.5% of its own norm over 93 steps". That page is one of the older built pages and is superseded by the blog page under the wiki's current-truth rule; both numbers are recorded in Superseded geometry claims. Note also that the monitor page's own attenuation figures differ from the current ones ("predicted 0.058, observed 0.056, agreement to 5%", against the current 0.0250 predicted / 0.0265 observed, 6%) - it is measuring a different quantity and should not be quoted as current.

Stage two: about 10%

PHASE3_VERDICT.md, addendum 2026-09-05:

It is not the A subspaces: seed-0 and seed-1 stage-2 A row spaces overlap at 0.022 (random), and stage-2 A drifts 10% from init (stage 1: 1.5%).

So the assumption that makes the stage-one chart legible is much weaker in stage two - 10% against 1.5%, as the verdict states them. It uses that to explain (without resolving) why the stage-two cross-seed attenuation slope is 0.116 rather than the r/d value of 0.025: "The extra cross-seed signal sits in a component every stage-2 adapter shares (different-factor floor 0.014 vs 0.002), consistent with B loading onto the shared part of A's drift. Not resolved; noted." See Stage two at a second seed.

Provenance for the 10% is prose only. No file on disk stores a stage-two A drift measurement; PHASE3_VERDICT.md is the sole source. Flagged in the section report.

adapter_effect.json

qwen35/analysis/adapter_effect.json is a 100-entry list, one per labelled trait, with fields trait, factor, keyed, sim_base, sim_s1, rep, leak, chars. The first entry is {"trait": "high_strung", "factor": "EmotionalStability", "keyed": "-", "sim_base": 0.1637198872731003, "sim_s1": 0.24038329403709985, "rep": 0.06592394902350215, "leak": 0.0, "chars": 892.0833333333334}.

The field names describe a behavioural measurement - a similarity of generated text to the trait under the base model versus under the stage-one adapter, a repetition rate, a leak count and a mean response length - not a weight-space one. It is included on this page only because the task brief filed it here; it is not a geometry result and nothing in the current blog page reads it.

No producing script exists in the repo and no built page found reads it, so what each field means is inferred from the names alone. Do not quote these numbers in the blog post without finding the code. Flagged in the section report; the behaviour agent should take it if it belongs anywhere.

Related: Weight-space geometry of the 134-adapter zoo, The seed floor, Stage-two geometry, Stage two at a second seed, Superseded geometry claims, Zoo training recipe.