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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

The distil page and its withdrawal

qwen35/distil_page was a findings page about where personality lives in the weights; it was withdrawn on 2026-09-01 for a units error in its headline claim and for resting on the invalid 200-token steering corpus, and replaced by the findings and monitor pages.

What "distil" refers to

Not a distillation experiment. distil is the directory name of a built HTML page whose title was "Where personality lives in a model's weights" (qwen35/distil_page/index.html.withdrawn-backup, distil_page/template.html). The word comes from the training recipe: qwen35/train_qwen35.py:38 notes "LR 5e-5 and DPO beta 0.1 are Open Character Training's published distillation" hyperparameters, and the page was about what those distillation-stage adapters contain.

It did not compare a prompted model against an OCT-trained LoRA. That comparison exists elsewhere in the project — the blog page's "Prompting versus training" section, built from the activation-space work — and belongs to Activation-space analyses: overview.

What the page contained

Its section headings, from the preserved backup:

1. Where personality lives in a model's weights
2. Five things that surprised us
3. Personality is everywhere, redundantly
4. The persona adapters contain the DPO adapter twice
5. The structure hides in the smallest directions
6. What each principal component is
7. What each factor is
8. Steering: does moving along a direction move behaviour?
9. Two claims we had to withdraw

Its data file qwen35/analysis/distil_data.json is still on disk (49 KB, 2026-08-29 17:00) with keys holo, deflate, copy, steer, factors_judged, pcs, fa, surprises, corrections.

distil_data.json#steer is the only surviving store of the withdrawn 200-token judged steering curves — nine directions, seven alphas, five scales. For example distil_data.json#steer[0] (PC1) gives Conscientiousness 3.739/5.042/5.042/5.5/5.542/5.458/5.5 across alphas -4/-2/-1/0/+1/+2/+4. These numbers should not be quoted as results; see The thinking-default trap and what it withdrew for why and Steering the base model along weight-space directions for what replaced them.

Why it was withdrawn

The current qwen35/distil_page/index.html is a 4.9 KB withdrawal notice dated Withdrawn 2026-09-01. It gives two reasons.

Reason 1 — a units error in the project's own analysis code.

"The persona adapters contain the DPO adapter twice." This was the page's second heading and had its own figure — a hundred dots landing on exactly 1.00.

It was a units error in our own analysis code. The sketch never read alpha/r from each adapter's config, and the merged personas carry a different scaling from their sources, so a factor of two appeared that was ours rather than the model's. The DPO strength in those adapters is correct.

Chasing that error did turn up a real defect in the merge, but a smaller and already-known one: PEFT's combination_type="linear" adds a cross term that Persona Cartography had previously identified and published. — qwen35/distil_page/index.html

The retracted claim is still in the data file: distil_data.json#surprises[0] reads "The merged persona adapters contain the DPO adapter twice ... proj(persona - stage-1, unit stage-1)/||stage-1|| = 1.001 +/- 0.015 across 100 traits; mismatched control 0.089 +/- 0.175." Treat that entry as withdrawn.

The merge investigation it triggered is recorded in qwen35/fix_persona_merge.py, qwen35/analysis/persona_merge_audit.json and qwen35/analysis/merge_audit.json, with the run log at phase10_runs/fixmerge.log (zoo-fixmerge.service, 2026-08-29).

Reason 2 — the steering panels rested on the invalid corpus. The full passage is quoted on The thinking-default trap and what it withdrew.

What replaced it

The notice names its replacements:

• What the Weights Know — qwen35/findings_page/index.html, built by qwen35/build_findings_page.py, whose own docstring says "Supersedes distil_page/, which carries a claim since shown to be wrong. That claim is not deleted here -- it is restated, corrected, and kept, because a findings page that quietly drops its retractions is less trustworthy than one that shows them."
• Personality Does Not Travel — qwen35/monitor_page/index.html, built by qwen35/build_monitor_page.py: "six tests of whether personality can be monitored from weights, with every criterion written before its test ran".
• "Twenty-two per-direction pages, one for each direction in the study, all built on the corrected corpus with per-alpha damage bars and verbatim quotes" — qwen35/direction_pages/, see Steering the base model along weight-space directions.

Note that build_findings_page.py still reads analysis/distil_data.json for deflate, pcs and fa — the withdrawal applies to two named claims and to the steering panels, not to the whole file.

The two corrections carried forward

qwen35/analysis/distil_data.json#corrections holds both retractions in the project's own words:

1. 'Stage 2 re-encodes and amplifies each trait's own direction' — The evidence was cos(SFT increment, own DPO) = +0.225 vs +0.015 mismatched. That gap is entirely the deterministic 2x copy above. After removing it the correlation is +0.0002.

2. 'Stage 2 sharpens factor clustering, 0.357 -> 0.548' — Stage-1 includes 34 Lexicon traits that carry no factor label to match, which drags its figure down. On the same 100 traits it is 0.508 vs 0.548 — a 4-point gain, not 19. At k=1 the direction reverses.

The first of these supersedes the H2 verdict detail in qwen35/analysis/live_results.json#hyp[1]; see Judged evaluations of the trait adapters.

Why the page still exists

The URL was shared. Deleting it would turn a correction into a dead link, and anyone holding the old link would have no way to learn that what they read was wrong. A withdrawal notice is more use than a 404, and considerably more use than the original. — qwen35/distil_page/index.html

The notice links two replacement artifacts by claude.ai/code/artifact/ URL rather than by local path.

Not to be confused with

qwen35/build_blog_corpus.py is unrelated to this page: it packs the corrected steering generations into analysis/blog_corpus.json for the current blog page's interactive explorer, keeping only alphas -2, -1, +1, +2 because "At |alpha| = 4 every direction loops, so those generations are evidence about breakage rather than about personality" (qwen35/build_blog_corpus.py).

Related: The thinking-default trap and what it withdrew, Steering the base model along weight-space directions, Judged evaluations of the trait adapters, built-pages-inventory, Activation-space analyses: overview, Glossary.