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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Stage two at a second seed

Fifteen traits put through the whole two-stage pipeline again from a second initialisation each find their own seed-zero stage-two adapter among 134 (15/15), the two seeds' cosine matrices agree at r = 0.978, and the attenuation slope is 0.116 - 4.6x the r/d prediction, unexplained.

What was run

PHASE3_VERDICT.md, addendum 2026-09-05, "stage 2 at a second seed":

Full OCT stage 2 (reflection + interaction generation, SFT, 0.25 merge) re-run for 15 traits (3 per factor, mixed keying) on the matched seed-1 stage-1 adapters, SFT seed 1, outputs under /oct/seed1. Nothing skipped; 313-375 optimizer steps; $226 planned, came in under.

The 15 traits, from qwen35/phase10_runs/results_oct2_15traits_...json#traits: helpful, cold, harsh; organized, disorganized, careful; relaxed, anxious, fretful; extraverted, quiet, assertive; intellectual, simple, unimaginative. Every sft record in that file carries sft_seed = 1 and oct_root = "/oct/seed1", and optimizer_steps ranges 313 to 375 - matching the addendum. .garden/journal/2026-09-05.md records the launch (Samuel chose the 15 traits; plan $226.28 against a $240 budget) and the completion ("all 15, under budget").

The cross-Gram compares the 15 seed-1 SFT LoRAs against the 134 seed-0 SFT LoRAs - not the merged personas (results/cross_gram_full_loras_introspection_x_seed1_loras_introspection.npz, names_a length 134, names_b length 15). The within-run block used for the attenuation regression is PC_WITHIN = the seed-0 stage-2 134x134 block.

The numbers

qwen35/analysis/crossseed_arms_stage2.json (a one-element list), against The seed floor's stage-1 matched arm:

The addendum's table quotes the same values rounded (same trait +0.0672 sd 0.0163; same factor same key +0.0195; different factor +0.0141; top-1 15/15; slope 0.116 Pearson 0.931).

One property of the stage-one result does not carry over, and the addendum does not mention it. In stage one the same-trait block is perfectly separated from the cross-trait block: min_same (0.01593) exceeds max_off (0.01584). In stage two it does not: min_same = 0.04248 is below max_off = 0.05141, and sep = -0.00892. Top-1 identification is still 15/15 and mean rank is still 1.0, so every trait wins its own column; but there is no longer a global threshold separating same-trait from different-trait pairs. Recorded in the section report.

The slope that is not r/d

PHASE3_VERDICT.md:

The slope is 4.6x the r/d prediction. 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%). 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.

The fact that makes the account plausible is in the table above: the different-factor floor is +0.014086716786298033 in stage two against +0.0018307002389548663 in stage one, while the same-trait mean rises only from +0.01806099632415457 to +0.0672081002737389. So a larger part of the stage-two cross-seed cosine is a component every adapter shares. The status of the explanation is explicitly unresolved.

The arrangement replicates

PHASE3_VERDICT.md:

Geometry: the seed-1 15x15 cosine block vs the seed-0 block for the same traits correlates at r = 0.978 raw, 0.987 trait-centred. Stage-2 arrangement replicates across seeds essentially perfectly on this sample.

The two blocks are results/cross_gram_full_seed1_loras_introspection_x_seed1_loras_introspection.npz (15 x 15, seed 1) and the corresponding 15-trait submatrix of the seed-0 stage-2 Gram. The blog page rounds this to "the two seeds' cosine matrices over those traits agree at r = 0.98". Fifteen traits is 105 off-diagonal pairs, so this is a small sample and the addendum says "on this sample".

.garden/journal/2026-09-05.md records the same numbers on the day ("same-trait 0.067, 15/15 top-1 of 134, slope 0.116 (not r/d), geometry across seeds r 0.98") and one operational note: cross_gram_full_on_modal.py choked on a nested subdirectory appearing in the output filename and had to be flattened.

Caveat that carries over from stage one

Same-trait seed pairs train on the same constitution and, in stage one, byte-identical corpora. Stage two regenerates its own transcripts per run, so the 15 seed-1 adapters did not see the same SFT rows as the seed-0 ones - the generation stage ran again. That makes the stage-two replication a slightly stronger claim about the pipeline than the stage-one one, and a weaker one about any particular corpus. No source states this explicitly; it follows from the run record (#stages.gen has 45 generation jobs in the seed-1 file) and is flagged as an inference rather than a quoted finding.

What the second seed later settled

PHASE3_VERDICT.md left the anomalous slope "not resolved", noting it was "consistent with B loading onto the shared part of A's drift". On 2026-09-08 the shared component was measured separately inside each seed's own LoRA-A frame on these same 15 traits: 0.2017 of the squared norm at seed 0 and 0.2000 at seed 1, cosine to the grand mean 0.4497 +- 0.0319 and 0.4475 +- 0.0330 (qwen35/analysis/stage2_frame.json#shared_component). The shared component is learned by the recipe, not inherited from the initialisation; what the frame fixes is only its coordinates, which is why the same learned direction in two different row spaces shows up as a cross-seed cosine of 0.014. See Stage two, explored.

Related: Stage-two geometry, The seed floor, Cross-seed geometry, Adapter effect and LoRA-A drift, Stage two, explored, Open Character Training (Maiya et al., 2025).