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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Superseded geometry claims

Every place the project reversed itself on the weight-space geometry, with dates and both versions - the trait-level withdrawal and its reversal, the bar in the wrong units, the objective mismatch, the scree source and the stale numbers still on disk.

Each entry gives the earlier version, the later version, the date of the reversal, and where both live. Nothing here is resolved by preference; where two files still disagree, both numbers are quoted.

1. The trait-level withdrawal, and its reversal

Earlier (2026-08-23, qwen35/PHASE3_VERDICT.md body).

The trait-level claim is withdrawn and the factor-level claim is upheld - individual trait directions are an artefact of the LoRA initialisation ... So the trait-level framing is withdrawn. There is no evidence here that an individual trait has a characteristic direction in weight space.

The preregistered bar was an across-seed self-cosine clearly above +0.2447, with an abandon line at 0.35; measured median was +0.0167.

Later (2026-08-22 22:40 UTC, addendum "THE TRAIT-LEVEL WITHDRAWAL IS REVERSED"). Note the addendum carries an earlier timestamp than the verdict body it reverses; the body was written 2026-08-23 and the addendum's own text says it reverses it.

Reinstated. A trait has a reproducible direction in weight space. It survives independent initialisation; the cosine is small only because the LoRA parameterisation confines each run to a different random rank-64 subspace, and 0.0264 is the overlap of two such subspaces, not a measure of signal loss.

Evidence: perfect separation (min same-trait 0.01509 > max different-trait 0.01490 across 5,320 comparisons), 40/40 top-1 identification, and the attenuation slope 0.0264 against the predicted r/d = 64/2560 = 0.0250. See The seed floor.

What is not reinstated, in the same addendum: "Same-trait pairs train on a byte-identical corpus, so what reproduces is that trait's training signal", and "1.7% alignment is statistically unambiguous and practically worthless as a transferable vector."

1a. And then it stands again, for nine days

The reversal was not the end of it. Two days after the addendum that reinstated the trait-level claim, the 2026-08-24 addendum partly re-withdrew it:

• The cross-seed self-cosine (+0.0167) is a seed-plus-objective floor, not a pure seed floor. The trait-level withdrawal STANDS (the preregistration bound it to this measurement as run), but the mechanistic reading "the init frame alone carries the magnitude" needs the 40 seed-B adapters retrained at the matched objective (~$17 at measured rates) before it is clean.

The 2026-09-03 addendum then discharges the caveat - "The 2026-08-24 caveat on attributing the floor to 'seed alone' is discharged; the page now quotes 0.0181" - without restating the reinstatement in words, and the blog page, which is current truth, describes trait identity as surviving the seed change ("across 40 traits retrained at a second seed, the matched adapter is the nearest neighbour among all 134 candidates 40 times out of 40 ... It is the initialisation and nothing else").

So the document holds four positions in sequence on one claim:

Note the two August entries are in the order the document prints them, not in timestamp order: the addendum carries the earlier timestamp and the verdict body it reverses is dated the next day.

Nothing in the file says "the withdrawal is reversed again". Anyone quoting a trait-level position must say which of the four they mean and date it.

2. "The preregistered bar was in the wrong units"

2026-08-22 addendum.

The floor compared a cross-basis cosine (0.0167) against a bar computed within a single basis (+0.2447, the median same-factor same-keyed different-trait pair in the seed-0 Gram). Those two numbers are not commensurable. Two independent rank-64 inits span near-orthogonal slices of a 2560-dim space, so a trait direction that reproduced perfectly modulo the subspace would still have scored ~0.02. The test could not have passed under any hypothesis.

The stated lesson: "A preregistered threshold needs its own check: compute what the statistic would read under the maximal version of the hypothesis, before committing to the threshold."

A second, smaller units problem sits inside the same bar and is unresolved. qwen35/PREREGISTRATION_phase3.md calls +0.2447 "the median cosine of same-factor, same-keying, different-trait pairs". The value 0.24472701836565558 is what qwen35/results/decomposition.json#test2.raw.same_polarity stores, and that key is a mean over 932 pairs. qwen35/results/compare_nulls_output.txt separately reports the median of the same class as +0.2467 (n = 932), and +0.2373 on the 40 seed-twinned traits (n = 130). Both are recorded; neither is preferred here.

3. The downstream flag about phase 7 steering

Earlier (PHASE3_VERDICT body). Phase 7 steering "assumed transferable trait directions and needs re-specifying before it is worth running."

Later (2026-08-22 addendum, "Correction to this document's own downstream flag").

That was wrong on two counts. The directions do transfer. And phase 7 never needed them to: it applies a concrete weight delta, found in the seed-0 basis, to the frozen base model, with matched random-direction controls at every dose. That operation is well defined regardless of how the direction was found. Phase 7 is runnable as specified.

4. compare_nulls.py's stored verdict

Earlier (qwen35/results/compare_nulls_output.txt, still on disk, never regenerated).

VERDICT: NO -- a trait trained twice looks NO MORE ALIKE than two different traits of the same factor and keying (+0.0167 vs +0.2467). The geometry is seed noise. EVERY EFFECT ABOVE IS VOID: TEST 1B, TEST 2 and TEST 2B are differences of a few hundredths computed inside a run-to-run wobble at least as large

Later (PHASE3_VERDICT, "Where I depart from the tooling").

That conclusion is wrong, and I wrote the check that produced it. It conflates two different things: the coordinates failing to correspond across bases, and the organisation failing to replicate. A cross-run cosine of 0.017 establishes the first. It cannot speak to the second, because structure can reproduce in a rotated basis while every corresponding vector reads as orthogonal. The cross-Gram is the wrong instrument for the question the verdict line was answering.

The file also predates the matched-objective retrains, so its REAL row is the only current line in it. Treat the whole file as historical.

5. The objective mismatch in every control arm

Discovered 2026-08-24 (PHASE3_VERDICT addendum), by comparing against the now-public Persona Cartography reference code.

the null launcher carried the corrected LoRA scale but not the loss configuration, so all 240 control adapters (shuffled, permuted, seed-paired) trained under plain sigmoid DPO -- loss_type ["sigmoid"], kl_coef 0.0 -- while all 134 sweep adapters trained with loss_type ["sigmoid","sft"], kl_coef 0.001.

Three claims were caveated: the cross-seed self-cosine became "a seed-plus- objective floor, not a pure seed floor"; the null table compared arms differing in objective as well as treatment; and the seed-B factor replication's attribution to "seed alone" inherited the caveat.

Discharged in two stages.

• 2026-09-03: the 40 seed-paired traits retrained at the matched objective. Same-trait cosine +0.01659 -> +0.01806; slope 0.0264 -> 0.0265; 40/40 top-1 both ways. The objective in isolation (same seed, same data) gives same-trait cosine +0.954. "The 2026-08-24 caveat on attributing the floor to 'seed alone' is discharged; the page now quotes 0.0181."
• 2026-09-05: shuffled and permuted retrained at the matched objective. TEST 1B residual +0.0001 -> +0.0001 and +0.0075 -> +0.0080; scree counts 11 and 0 unchanged. "The objective mismatch changed nothing to three decimal places."

Tooling change: cross_gram_on_modal.py now refuses matched pairs whose runmeta disagree on loss_type / loss_weights / kl_coef.

See The null control arms and The seed floor.

6. Two null training runs that were thrown away

Not a claim reversal but a record of what the logs are.

• .RSLORA-INVALID: the trainer's default said rsLoRA while the sweep used plain LoRA, so 240 control adapters were trained at scale 16.0 against the sweep's 2.0 and the whole phase was rerun (qwen35/launch_nulls.sh, the comment above export PC_USE_RSLORA=0).
• .FAILED-datarace: the trainer's corpus-content gate at train_qwen35.py:1050 refused to train when the sha256 of the jsonl inside the container disagreed with the corpus the driver declared - "a real-vs-null corpus swap or a stale upload -- the two are indistinguishable by prompts alone. BLOCKER."

7. The elbow

Earlier: with only the elbow to go on, six components looked like a generous description of the cloud. Later (blog page): against the structureless null the real spectrum stays above the floor through eleven components, "which happens to sit right beside the twelve the factor side retains by an entirely separate route. So describing six components is conservative, not generous. I had that backwards when I only had the elbow to go on."

8. The stored scree real curve

.garden/journal/2026-09-05.md:

Scree null regenerated from the matched Grams; stored real came from gram_sketch100.npz, not gram_sweep (took a minute to find).

Found by a reproduction check. What is still observable on disk:

• #real is byte-for-byte the same in analysis/scree_null.json (2026-09-02 20:20) and analysis/scree_null_matched.json (2026-09-05 12:15) - 13.36216610744329, 11.207049613198226, 5.789132715752... - so it was not recomputed when the null arms were.
• Those values equal analysis/geometry_stage1.json#explained_var_top10 (0.134, 0.1122, 0.0577, ...) after rounding, and that file is computed over the 100 labelled traits from the stage1_k32 sketches.
• The exact 134-trait double-centred spectrum is different: results/fa_qwen35.json#pca_from_gram.centered_var_pct = 12.4946, 10.9847, 4.9679, ...

So the real curve on the current page is a 100-trait sketch-derived spectrum. Nothing on disk shows it being regenerated from results/gram_sweep.npz. The sketch fidelity (r = 0.9996) makes a shape change unlikely, but the provenance is what it is. See PCA, the scree curve and the two nulls.

9. Stale numbers still sitting in current data files

• qwen35/analysis/pc_loadings.json#seed.same_trait_cos = 0.01658715905967511 and #seed.same_trait_deg = 89.04958220635345 are the original, objective-mismatched seed-paired figures. The blog page's prose quotes the matched 0.018 while its own data file carries 0.0166. The 89-degree figure the page prints was computed from the stale cosine.
• qwen35/analysis/scree_null.json#perm_mean_abs_diff = 0.1798544716598768 is what the blog page's prose quotes ("0.18 percentage points apart on average"), while the file the figure is drawn from, scree_null_matched.json#perm_mean_abs_diff, is 0.10336514494213549.

10. Contradictions between files, unresolved

Coverage at k = 2. analysis/viz.json#coverage.2.gap_deg = 3.9655382092691895 (z = +1.5803426589637861, nearest introverted) against analysis/alien.json#k_sweep.2.gap_deg = 0.9282998475514116 (z = -2.280849861251385, nearest unsympathetic). The two agree to within a fraction of a degree at every other k. Opposite signs mean opposite conclusions about whether the 134 words cover the plane better than chance. The blog page uses alien.json, which was written 46 minutes later on the same day. See UMAP, sphere and other layouts.

LoRA-A drift. analysis/align_summary.json#a_drift = 0.014605041334818797 (1.5%, quoted by the blog page) against qwen35/build_monitor_page.py line 112: "it drifts 4.5% of its own norm over 93 steps". The monitor page is an older built page; the blog page's 1.5% is the figure the current draft and the analysis files carry. The same monitor page also quotes the attenuation as "predicted 0.058, observed 0.056, agreement to 5%", where the current figures are 0.0250 predicted / 0.0265 observed at 6% - a different quantity that should not be quoted as current. See Adapter effect and LoRA-A drift.

Sketch fidelity. qwen35/build_blog_page.py line 1584 (prose): r = 0.9996. qwen35/build_monitor_page.py line 209: "independently computed exact Gram at cosine 0.99944". No output file storing either was found on disk; analysis/validate_100.json is training metadata, not sketch validation. See Weight-space geometry of the 134-adapter zoo.

Cross-seed category counts. analysis/crossseed_arms.json#[0] records same_factor_same_key n = 368, same_factor_opp_key n = 392, diff_factor n = 4560 (which with the 40 same-trait pairs totals 5,360, while #[0].n_pairs is 5,320). The 2026-08-22 addendum's table gives n = 358, 386 and 4336 for the same three classes. The means agree (+0.00593 / -0.00271 / +0.00133); only the counts differ. Unexplained.

Permutation p values. results/decomposition.json reports p = 4.999750012499375e-05 throughout (20,000 permutations); results/compare_nulls_output.txt prints the same results as 0.00005; the PHASE3_VERDICT replication table prints 0.00010. All three describe the same permutation floor at different resolutions.

Centroid deflation variance removed. analysis/polarity_deflation.json#stage1.var_removed = 0.045566670770937776, while analysis/manifold_ideas.md and qwen35/build_findings_page.py both quote "5.3% of variance" for the same experiment. See Polarity, bipolarity and the trait graph.

11. Numbers whose provenance is prose only

Recorded so that a blog post copying them knows what it is copying.

Not on this page

~/projects/persona-curvature/CONTEXT.md (written 2026-08-14) carries its own explicit SUPERSEDED markers - the explanation for why weight-space control failed, and the oracle-direction yardstick - but they are about the pre-zoo model organism (Qwen2.5-3B, sycophancy) and not about the zoo's geometry. They belong to the history section.

Related: The seed floor, The null control arms, PCA, the scree curve and the two nulls, Cross-seed geometry, Adapter effect and LoRA-A drift, UMAP, sphere and other layouts, Weight-space geometry of the 134-adapter zoo.