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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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Superseded claims

Every claim the project made and later reversed or corrected, dated, with the value as first stated, the value that replaced it, and the page that holds the current version; opens with the sentences on the live blog page that are wrong as printed.

This is the project's record of its own reversals: every claim that was stated and then corrected, withdrawn or replaced, across the whole project. Each entry gives the claim as first stated (value, where, when), what replaced it (value, where, when), and the wiki page that holds the current version.

Two companion pages. Source contradictions holds disagreements where neither side is known to supersede the other. Open questions holds what is unconfirmed, never run, or still a decision. The geometry section keeps its own detailed record at Superseded geometry claims; section 3 below is a pointer list into it rather than a second copy.

Two conventions. Numbers are quoted verbatim from the file that stores them, at that file's precision. Where a claim reversed more than once, every position is listed with its date, because a reader quoting "the project's position" must say which one.

1. Blog page statements that needed changing (the page is historical since 2026-09-15)

qwen35/blog_page/index.html, built by qwen35/build_blog_page.py, was the authoritative current state of the results on 2026-09-07. From 2026-09-15 the current statement was qwen35/POST_DRAFT.md (post-draft) with the analysis files it cites, and since 2026-10-30 it is the post Persona Geology (Post: Persona Geology (2026-10-30)); the blog page is a historical built page. The items below are kept as the record; where the post draft fixed one, the note says so. L1: fixed, the draft says "the rejected side being a character at the opposite pole". L2 and L3: moot, the draft quotes the matched null file and the matched seed floor 0.018 with no angle. L4: the draft claims only adapters, factor adapters and stage-two transcripts on the Hub; the preference data is part of the planned results release (Code and data map). L7: fixed, Appendix A0 states the 6/14 keying. L9 and L10: the draft quotes 0.9999992 for the finite-difference check and does not quote the sketch-validation figure. L13: the PC pages are no longer primary. They are separated by what a writer had to do about each.

1a. Wrong as printed - change the sentence

1b. Correct but needs a qualifying clause

1c. Printed but not backed by a file

These are not known to be wrong. They are on the page as prose or as hard-coded strings, with no output file storing them, so a blog post copying them is copying a builder script.

1d. Published Hugging Face cards that are stale

Not the blog page, but public and wrong, and flagged as owed work by qwen35/POST-BATCH3-TODO.md item 4 and qwen35/analysis/hf_dataset_audit.md section 5.

• qwen35/zoo_page/MODEL_CARD.md: "45 adapters", the "0.886 +/- 0.111 across all 45" band and the "14 of 45" row-drop statistic. The zoo is 134.
• qwen35/zoo_page/DATASET_CARD.md: "51 constitutions", "194 of 204 files", size_categories: 100K<n<1M, and "so sft_data is often smaller than 12,000" - hf_dataset_audit.md parsed four files at exactly 12,000 rows, including two the card names as heavy losers, and concludes the drop happens at the SFT tokenizer, which is what MODEL_CARD.md says.

Current pages: Hugging Face artefacts, Stage two - introspection SFT and the 0.25 persona merge.

2. Zoo construction

Z1. rsLoRA on, then off. First stated: qwen35/plan.json#defaults.use_rslora = true, with #defaults.rslora_note calling rsLoRA the default and plain LoRA an ablation arm. qwen35/paper_notes.md section 3.2 (2026-08-19) then argued at length against the project's own rsLoRA setting, calling it the largest silent divergence from both source papers. Replaced by: all 134 records in qwen35/phase2_runs/archive/phase5_sweep_134.json read use_rslora: false, expected_scaling: 2.0; qwen35/RUNNER_TASK.md says "plain LoRA alpha 128 rank 64 (use_rslora False)"; qwen35/PHASE3_VERDICT.md's 2026-09-03 addendum says "rslora off". Current: Stage-one training configuration, Zoo training recipe.

Z2. Pair volume and hyperparameters in the plan. First stated: qwen35/plan.json phase 1 "214 pairs per trait" and phase 2 "rank 32 / LR 5e-6", both flagged in paper_notes.md section 3.11. Replaced by: 500 pairs targeted, 445 trained, at rank 64 and LR 5e-5. Current: The shared prompt pool and the 445-prompt intersection, Stage-one training configuration.

Z3. The stage-two adapter count, six times. Dates each figure was correct: 41 (qwen35/upload_adapters.py docstring), 45 (qwen35/zoo_page/MODEL_CARD.md, twice), 51 and 61 (qwen35/POST-BATCH3-TODO.md), 103 (qwen35/upload_zoo_batched.py docstring, which gives stage2_introspection/ 103, persona_merged/ 103, persona_exact/ 103 against stage1_dpo/ 134), 119 (qwen35/analysis/hf_dataset_audit.md), 134 (qwen35/analysis/merge_audit.json, build_blog_page.py, and the OCT-2 results files: 100 primary plus 34 lexicon). Current: 134. See Stage two - introspection SFT and the 0.25 persona merge, Hugging Face artefacts. The two published cards are the stale ones (section 1d).

Z4. Transcript repository coverage. First stated: qwen35/analysis/hf_dataset_audit.md (2026-08-29): 50 of 134 traits on the repo, 194 of 536 files, one trait quarantined and two partial. Replaced by: qwen35/phase10_runs/upload_datasets.log (2026-08-30 10:30): "COMPLETE: all 536 expected files present... nothing quarantined". The audit file is not annotated to say so. Current: Hugging Face artefacts.

Z5. "Nothing described here is published." First stated: qwen35/PENDING-CONTENT-REVIEW.md. Replaced by: the same 2026-08-30 upload log, which shows the previously quarantined rows uploaded as "previously reviewed and cleared", plus 60 written clearance decisions in qwen35/phase10_runs/adjudications.json. Whether that reflects a decision by Samuel is not recorded in any file; see Open questions. Current: Hugging Face artefacts, zoo-build-governance.

Z6. The spend ledger: $195.90, then ~$370.65, then the meter. First stated: $195.90 realised, being the sum of qwen35/plan.json's closed phases[*].actual values as of 2026-08-23 00:05, quoted against a $1000 ceiling. Replaced by (2026-08-23 00:05): ~$370.65 measured - qwen35/plan.json#ledger_defect_note gives "Modal $308.83 + OpenRouter $61.82", and names three mechanisms for the $174.75 gap (billed is not incurred; the app-name collision that billed the 134-run sweep under phase 2; a phase costs its failures - phase 3 recorded $111.04 against $206.84 measured because of the .FAILED-datarace and .RSLORA-INVALID predecessors). Superseded again by the container-minute meter: $2,240.52 of $2,400, the standing reading in qwen35/phase10_runs/zoo40_meter.log through 2026-09-07T16:25:05Z (first written at line 25572, 2026-09-05T21:43:37Z). It is a meter, not a billing read: PRIOR=46.90 plus integrated GPU-hours at a flat $2.10/hour from polling modal container list, pricing CPU containers at the A100 rate. It is live and has started moving again since the full-OCT replication launched on 2026-09-07 (see Open questions); quote it with its read time. Note on the file: plan.json's ledger_defect_note still says "Sum of actual = $195.90", but the file's own phases[3].actual now reads 206.885 where the note describes phase 3 as $111.04 recorded. The $195.90 sentence no longer describes the file it sits in. Current: zoo-spend-ledger, costs.

Z7. The budget ceiling itself moved three times. $200 (CONTEXT.md section 6, "against the $200 set") -> $1000 (qwen35/plan.json#budget) -> $2400 (qwen35/zoo40_meter.sh, BUDGET=2400.00, "raised 2026-08-29 against a $5k grant"). qwen35/HANDOVER.md still frames spend against $1000 and was not updated. Anyone quoting "of the budget" without a date is quoting one of three denominators. Current: costs, zoo-spend-ledger.

Z8. The bare "$195.90 actual against a $1000 ceiling" was deleted, not annotated. qwen35/HANDOVER.md commit f2e02b4 removes the sentence and replaces it with a SUPERSEDED banner, on the stated principle that "a caveat does not travel with a number". Current: costs, lesson-a-phase-costs-its-failures.

Z9. The enumerated constitution anchor. First stated: every constitution ended with an anchor paragraph that named one Goldberg marker per Big Five factor. Replaced by (2026-08-19): a generic anchor block, because naming the markers could have manufactured the Big Five result under test. The enumerated form was retained as an ablation arm and, as far as any file shows, never trained (see Open questions). Current: The anchor block and its 2026-08-19 revision.

Z10. The secondary trait source. First stated: draws 1 and 2 sampled Allport and Odbert's 1936 list; draw 1 had no semantic constraint and returned 14 of 40 unusable words, draw 2 added WordNet-gloss filters and was worse. Replaced by: draw 3 from Condon et al.'s trait-descriptive set, 5,636 to 2,303 to a 40-cluster k-means draw. Both discarded files are still on disk as qwen35/traits_secondary_DISCARDED_draw1.json and ..._draw2.json. Current: The 34 Lexicon traits and how they were drawn, The two discarded secondary draws.

Z11. The persona merge. First stated: the published personas were assembled with PEFT's linear adapter combination. Replaced by: PEFT sums LoRA factors, not deltas, so the merged delta carries a cross term worth 0.795 to 0.823 of its Frobenius norm across all 134 traits (qwen35/analysis/merge_audit.json). An exact concatenation merge was measured and published alongside. Current: The persona merge and its correction.

3. Geometry

Superseded geometry claims is the full record, with both texts quoted for each reversal. One line each here, with the section number to open.

• The trait-level withdrawal, reversed and re-withdrawn and discharged - four dated positions on one claim (2026-08-22 addendum, 2026-08-23 body, 2026-08-24 addendum, 2026-09-03 addendum plus the blog page). Superseded geometry claims sections 1 and 1a. Current: The seed floor.
• The preregistered +0.2447 bar was in the wrong units and "could not have passed under any hypothesis". Section 2. Current: The seed floor, lesson-bar-in-the-wrong-units.
• The verdict's own downstream flag on phase 7 ("assumed transferable trait directions... needs re-specifying") - "That was wrong on two counts." Section 3. Current: Steering the base model along weight-space directions.
• qwen35/results/compare_nulls_output.txt's "EVERY EFFECT ABOVE IS VOID" is overridden by its own author and the file was never regenerated. Section 4. Current: The null control arms, lesson-coordinates-are-not-structure.
• The objective mismatch in all 240 control adapters, and the seed floor moving +0.01659 -> +0.01806 on the matched retrain (2026-09-03), with the null arms following on 2026-09-05. Section 5. Current: The seed floor, The null control arms, lesson-objective-must-travel.
• Two thrown-away null training runs, .RSLORA-INVALID and .FAILED-datarace. Section 6. Current: The null control arms.
• The elbow reading was backwards: six components is conservative, not generous. Section 7. Current: PCA, the scree curve and the two nulls, explainer-elbow-figure.
• The stored scree real curve is a 100-trait sketch spectrum and was never regenerated from results/gram_sweep.npz. Section 8. Current: PCA, the scree curve and the two nulls. It is also an open item - see Open questions.
• Stale numbers still sitting in current data files (pc_loadings.json#seed, scree_null.json#perm_mean_abs_diff). Section 9; they are also blog-page items L2 and L3 above.
• File-to-file contradictions the geometry page records without resolving. Section 10; each is written up with a settlement test in Source contradictions.
• Numbers whose provenance is prose only. Section 11; the blog-facing subset is section 1c above.

Three geometry reversals that page does not cover:

G1. PC4 named as a sycophancy axis. First stated: PC4 was described as a sycophancy axis. Replaced by: the blog page - "naming it after sycophancy was a mistake"; its highest loadings are timid, self-pitying, fearful, guilty. The replacement claim was then preregistered in qwen35/PREREG_alignment.md as prediction 1 and held. Current: PC4 - Affirmation and Self-Concern, Alignment-relevant traits in the geometry.

G2. qwen35/analysis/sphere_layout.json and qwen35/results/umap_embeddings.json. The sphere layout is superseded by sphere_page.json; the UMAP embeddings (2026-08-23) read the pre-matched null Grams and are historical. Current: UMAP, sphere and other layouts.

G3. The 2026-09-05 stage-2 addendum reports 15/15 top-1 and does not report the separation. Stage-one separation is positive and the 2026-08-22 addendum makes "perfect separation, no overlap at all" one of its three structural facts. At stage two qwen35/analysis/crossseed_arms_stage2.json#[0] has min_same 0.04248138800846624 below max_off 0.05140579106757132, giving sep -0.00892440305910508. So the separation property does not carry to stage two, and no document says so. Current: Stage two at a second seed; also an open question.

4. Behaviour

B1. The 200-token steering corpus. First stated: qwen35/phase10_runs/steer_results.json and judged_steer.json, plus the pre-aggregated curves preserved at qwen35/analysis/distil_data.json#steer. Replaced by (2026-08-29): steer_results_fix.json / judged_steerfix.json. Qwen3.5's chat template defaults thinking ON, so all 1,512 original generations were truncated reasoning preambles; the corpus was regenerated with enable_thinking=False at 512 tokens. Current: The thinking-default trap and what it withdrew, Steering the base model along weight-space directions, lesson-thinking-default-trap.

B1a. "Degeneration under steering in this project is overwhelmingly a negative-alpha phenomenon." First stated (2026-09-09): Fisher norms of the steering directions, from qwen35/analysis/fisher_norms.json#comparisons.sign_asymmetry_of_degeneration - over 22 directions with both signs, all 22 loop at least as much at alpha -2 as at +2 and 18 have a loop rate of exactly zero at +2. Qualified (2026-09-10): the observation is a fact about a fixed alpha grid, not about sign. At matched Fisher dose the mean loop rate is 0.075 on the suppressing sign and 0.1375 on the amplifying sign over ten directions, and 0.075 against 0.2167 over the five factors; the largest loop rate in that run is FA_Arousal's amplifying side at 0.625 (qwen35/analysis/matched_dose_steering.json#summary). Where a direction's amplifying sign is the flatter one, matching the dose pushes it to a larger alpha and it degenerates there. The original counts stand for the corpus they were measured on. Current: Suppression versus amplification at matched Fisher dose, Fisher norms of the steering directions.

B1b. The suppress-versus-amplify asymmetry as evidence for a mechanism. First stated: qwen35/POST_DRAFT.md and Steering the base model along weight-space directions - at alpha plus or minus 2, "four of five suppress harder than they amplify", offered alongside Gradient Atoms' reading that suppression is structurally easier (Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms). Corrected (2026-09-10): the asymmetry is headroom. Matching the Fisher dose takes the raw suppress/amplify ratio from 1.8957528957528955 to 1.6455696202531644 (8 of 10 directions to 6 of 10), and dividing each change by the room that existed in the direction it moved - delta / (7 - base) upward, delta / (base - 1) downward - takes it to 0.9261912176364205 published and 0.7946215813581846 at matched dose, i.e. amplification is the stronger move per unit of room in both arms (qwen35/analysis/matched_dose_steering.json#summary.all, #published_equal_alpha_summary.all). The base sits at 5.916666666666667 of 7 on Conscientiousness and 5.625 on Intellect in the run's own base condition, so there is four to five times as much room downward as upward. The published table is not wrong; it is not independent evidence for the mechanism story, and post-draft should not present it as if it were. Current: Suppression versus amplification at matched Fisher dose.

B1c. A published steering alpha is a full-strength perturbation. Implicit throughout: qwen35/steer_fix.py adds the increment into bf16 parameters and every published steering alpha was read as the intended dose. Measured (2026-09-10): the Frobenius norm survives (retention 0.9216 to 1.033505578742928, median 1.0027071769953744), but the KL does not: KL(bf16) / KL(exact) at the same nominal alpha runs 0.7117708405618217 to 0.9782241941431974 with a median of 0.9059646365364478 over 120 (direction, alpha) cells, and the shortfall depends on the alpha - the median is 0.9423571405594928 at |alpha| 2 and 0.7510758772738881 at |alpha| 1 (qwen35/analysis/matched_dose_steering.json#bf16_vs_exact, Suppression versus amplification at matched Fisher dose). Quantisation keeps the length and spends part of it on isotropic noise. The published alpha plus or minus 2 therefore delivered about 94 per cent of its intended dose and an alpha of 1 about 75 per cent. No published result reverses on this; the correction is to the unit, and it compounds with the ref correction recorded in the 2026-09-08 audit row of section 7 below ("alpha is measured in units of one trait adapter's worth of weight change"). Current: Suppression versus amplification at matched Fisher dose, Fisher norms of the steering directions.

B2. The zoo-steer2 and zoo-steer3 runs. Replaced by (2026-08-30): zoo-steerfix2 / zoo-steerfix3. No results file survives either of the originals. Current: Steering the base model along weight-space directions.

B3. qwen35/distil_page/index.html is withdrawn. First stated: a built page titled "Where personality lives in a model's weights" - module holography, the DPO-twice claim, variance deflation, PCs, FA and steering. Replaced by (2026-09-01): a withdrawal notice at the same URL, with the original preserved at .withdrawn-backup; its content is carried by findings_page, monitor_page and the 22 direction_pages. Current: The distil page and its withdrawal, built-pages-inventory.

B4. "The persona adapters contain the DPO adapter twice." First stated: qwen35/analysis/distil_data.json#surprises[0], with proj(persona - stage-1, unit stage-1)/||stage-1|| = 1.001 +/- 0.015 across 100 traits against a mismatched control of 0.089 +/- 0.175. Replaced by (2026-09-01): the distil_page withdrawal notice - "It was a units error in our own analysis code. The sketch never read alpha/r from each adapter's config... The DPO strength in those adapters is correct." The data file still carries the withdrawn claim. Current: The distil page and its withdrawal.

B5. Stage 2's relation to each trait's own direction. First stated (2026-08-29 14:19): qwen35/analysis/live_results.json#hyp[1].detail - "cos(SFT increment, DPO adapter) is +0.225 same-trait vs +0.015 mismatched, a 15x gap. Stage 2 re-encodes and amplifies each trait's own direction." Replaced by: qwen35/analysis/distil_data.json#corrections[0] - "That gap is entirely the deterministic 2x copy above. After removing it the correlation is +0.0002." live_results.json was never updated. Caution: the correction itself rests on the 2x copy, which B4 then withdrew as a units error. Only the residual-orthogonality figure (+0.0002) and the RSA structure claim at distil_data.json#surprises[2] should be carried forward. Current: The distil page and its withdrawal, Stage-two geometry.

B6. "Stage 2 sharpens factor clustering, 0.357 -> 0.548." Replaced by: qwen35/analysis/distil_data.json#corrections[1] - "0.508 vs 0.548 - a 4-point gain, not 19. At k=1 the direction reverses." Current: The distil page and its withdrawal, Stage-two geometry.

B7. qwen35/zoo_page/index.html is a 51-adapter readout (TOTAL = 51), superseded 2026-08-24 by qwen35/site_traits/ at 134. Current: built-pages-inventory.

B8. "13 of 24 AI-identity disclaimers belong to mean_assistant_axis." Replaced by: qwen35/analysis/qual_pc.json#directions[PC2].surprise - they belong to PC2's negative pole. On the corrected corpus the mean_assistant_axis negative side is flat at baseline (manual counts 8/7/8/7/4/3/1 across alpha -4/-2/-1/0/+1/+2/+4). Current: Qualitative reads and adjudications, PC2 - Hedging and Bluntness.

B9. PC1's named scale. First stated: qwen35/PREREG_steerfix.md registers PC1 as an Agreeableness claim - "PC1 - affect-centring vs procedural detachment... Agreeableness slide survives". Replaced by the measurement: qwen35/analysis/steerfix_replication.json#PC1.named is "Conscientiousness", and qwen35/analysis/qual_pc.json#directions[PC1].judged_profile repeats "named Conscientiousness". The replication file is current; the damage half of the prediction did replicate. Current: Steering the base model along weight-space directions, PC1 - Flooding and Composure.

B10. axis_Intellect "buys abstraction with Conscientiousness". First stated: qwen35/PREREG_steerfix.md - "largest range, buys abstraction with Conscientiousness... Intellect increasing, Conscientiousness decreasing." Replaced by: qwen35/analysis/qual_axes.json#directions[axis_Intellect].surprise - judged Conscientiousness on this axis rises with alpha (+0.20 slope, 2.54 at -4 to 5.83 at +1). "The debt in the corrected data runs the other way." Current: Steering the base model along weight-space directions, axis_Intellect (the named Intellect / Openness axis).

B11. The STEER134 coherence-collapse argument - downgraded, not superseded. First stated: qwen35/analysis/manifold_ideas.md builds a coherence-collapse argument on qwen35/results/steer134_judged.json (fa1 0.42/10 at +2, random1 8.5 at +2, pc1 9.25 at -1 against 2.0 at -2). Status now: that file has coverage.complete: false, with 11,010 of 28,664 units unjudged and many cells at n = 4-7 against n_expected 12, against only two random controls. The argument is marked unconfirmed on Steering the base model along weight-space directions. Nothing replaced it; it should not be cited as a finished result. See Open questions.

5. Activation space

A1. "Prompt wins 16 of 22." First stated: the first cross analysis, run on raw (uncentred) activation vectors, found the combined model nearer the prompt's persona than the adapter's in 16 of the 22 opposite-keying conflict pairs - that a system prompt largely overrides a trained trait. .garden/journal/2026-09-05.md, verbatim: "First analysis on raw vectors said 'prompt wins 16/22' - an artefact of the shared component (all pairwise cosines ~0.8)." Replaced by (2026-09-05): the trait-specific analysis gives 8 of 22 and the conclusion becomes "neither wins". qwen35/analysis/actspace_cross_geometry.json#resp_specific; qwen35/ACTSPACE_RESULTS.md; the blog paragraph is built from resp_specific (build_blog_page.py: X = J["resp_specific"]). Lesson filed the same day: any comparison of steered activations must remove the common shift first, or it measures "was the model steered" and not "toward what". Current: Adapter x constitution, 16 x 16, Prompting versus training.

A2. "A character English has no word for" - withdrawn. The activation-space transplant is part of why. Full entry at F5 below; current page Where no word goes - the hole and the alien direction.

A3. A stale docstring. qwen35/analyse_actspace_cross.py line 12 says "Prompt alone is 1.00 by construction, adapter alone was 0.62", but that script's own trait-specific block prints 0.70 (actspace_cross_geometry.json#resp_specific_adapter_alone = 0.7044131755828857). Source comment only; nothing downstream reads it. Current: Adapter x constitution, 16 x 16; the 0.62/0.70 pair is Source contradictions entry S2.

6. Literature notes

L-1. qwen35/paper_notes.md section 3 is a 2026-08-19 plan, and the run overtook at least two of its items. Anyone reading that section should be warned at the top.

• 3.2, rsLoRA. The note calls the planned rsLoRA the largest silent divergence: scaling 16.0 against both papers' 2.0, an 8x larger functional step, and warns that the project's Frobenius norms will land about 8x above Persona Cartography's 6.08-6.53 band, invalidating any comparison to their Table 1. The zoo turned rsLoRA off (Z1 above). The divergence is closed and the norms are comparable - which makes the missing norm table a cheap open item (see Open questions).
• 3.5, prompt volume. The note argues against a planned "500 pairs per trait, no general-prompt mix", giving roughly 16 optimizer steps at batch 32. The executed zoo ran 445 pairs and 13 optimizer steps on one pool shared across every trait. The argument survives and sharpens; the numbers in section 3.5 must not be quoted as the zoo's.

Current: Where the recipe diverges from the two source papers, Open Character Training (Maiya et al., 2025), Stage-one training configuration.

L-2. paper_notes.md section 4 lists eight things that could not be retrieved. Four are now retrievable (checked 2026-09-07):

Items 5, 6 and 7 were not pursued and are not load-bearing; item 8 (prior art doing factor analysis on weight deltas) is still not found. Current: Persona Cartography (Baines et al., 2026), Open Character Training (Maiya et al., 2025), Factor-analysis methods reference.

L-3. "LoRAcles" attributed to arXiv:2601.11207. Replaced by: De Schamphelaere, Bauer, Nanda, Ong, ICML 2026 MI workshop, OpenReview x9MbM7QmQN. Harness memory, 2026-08-16. Current: Origin, and how the question changed shape three times.

L-4. Persona Cartography's date. CONTEXT.md gives 10 Jul 2026, which is the LessWrong crosspost; arXiv v1 is 8 Jul 2026, 21:00:44 UTC. Current: Persona Cartography (Baines et al., 2026).

L-5. "Ashton & Lee 2004 HEXACO". Replaced by: the lexical-approach defence is Ashton & Lee 2005 (European Journal of Personality 19(1), 5-24); the HEXACO advantages paper is Ashton & Lee 2007 (PSPR 11(2), 150-166); the 2004 citation that circulates is Lee & Ashton 2004 (Multivariate Behavioral Research 39(2), 329-358), reversed author order. Current: Big Five reference bibliography.

L-6. "Cartography's modification is that the teacher picks the rejected response." Recorded in the harness reading list as a disagreement between two of its own extractions, unsettled. Settled 2026-09-07: both sides come from the teacher, one conditioned on the amplifier constitution and one on the suppressor; the polarity of the constitution determines the label. paper_notes.md section 2.3 quotes App A.1.1 verbatim, and vendor/persona-cartography/src/training/paired_dpo/pairing.py confirms the shape. Nobody picks anything. Current: Persona Cartography (Baines et al., 2026).

L-7. Big Five keying stated as 10/10 across all five factors in an early draft of Big Five reference bibliography. Replaced by: Emotional Stability is 6/14 (item L7 in section 1b). Current: Big Five reference bibliography, The 100 Goldberg primary traits.

7. Conversation-stated numbers

Numbers Samuel was told in chat that later data changed. Transcript references are to ~/.claude/projects/-home-vibe12-projects/981fa3b5-8a9b-405e-9e84-70cc615d9873.jsonl by line, as 981fa3b5:<line>; the extracts are in wiki/raw/.

F1. The seed floor: 0.0166 and 89.0 degrees. Stated: 981fa3b5:11041, 2026-09-02T20:35:17Z, in the reply to the external review. Undermined 29 minutes later: 981fa3b5:11177, 21:04:51Z, found that the 240 control adapters had trained under a different objective from the 134 (plain sigmoid DPO with kl_coef 0.0 against loss_type ["sigmoid","sft"] with kl_coef 0.001). Replaced by: +0.01806 matched against +0.01659 original, qwen35/PHASE3_VERDICT.md 2026-09-03 addendum (line 275). Current: The seed floor.

F2. The separation margin: 0.01509 / 0.01490. Stated: 981fa3b5:11177. Replaced by: 0.01593 / 0.01584 in the 2026-09-03 addendum - and then removed from the page entirely per 981fa3b5:12639. The string 0.01593 does not appear in qwen35/blog_page/index.html. Current: seed-paragraph-feedback, The seed floor.

F3. "PC1 is not a Big Five factor", and PC2 is "close to Agreeableness". Stated: on the page until 2026-09-02. Replaced by: 981fa3b5:11041 - PC1 is a blend, +0.82 Agreeableness / -0.70 Conscientiousness; PC2 is +0.81 Extraversion (the geometry, against the blind judge's reading of Agreeableness). Both corrections are on the live page. Current: external-review, PC1 - Flooding and Composure, PC2 - Hedging and Bluntness.

F4. Blog skeleton B: "PC1 is a common direction, the personality axis; the Big Five appear from PC2 on." Stated: 2026-09-04, 981fa3b5:12109 - two days after F3 and contradicting it, and contradicting 981fa3b5:12973, which says the personality axis is the mean adapter, reported separately, with the PCs describing variation around it. The live page agrees with the latter. Do not carry the skeleton B phrasing into the post. Current: blog-skeletons, PC1 - Flooding and Composure.

F5. "A character English has no word for" - withdrawn. Stated: on the page as the hole result's headline. Replaced in two steps: first by "the hole is in the sample, not the language" (981fa3b5:11094); then, after the three hole-word adapters and the activation-space transplant, by "suggestively, not convincingly, insouciant" (981fa3b5:13065). 981fa3b5:13036 calls the original "the most quotable line on the page and the one I trust least". The activation arm's own framing is "a direction the adapters leave open". Current: Where no word goes - the hole and the alien direction, external-review, most-interesting-claims.

F6. The elbow "after component 2" - withdrawn. Stated: 981fa3b5:10919. Replaced by: 981fa3b5:10939 showed four criteria give 1, 2 or 3; the figure now shows a null comparison instead and reads "about eleven components above noise". Current: explainer-elbow-figure, PCA, the scree curve and the two nulls.

F7. Using the permuted arm as a null the real spectrum must beat. Stated: implicitly, in how the null table was read. Replaced by: 981fa3b5:10963 - a category error. The permuted arm learns the same personas under wrong names, so its spectrum should match; the shuffled arm is the structureless baseline. Current: The null control arms, lesson-test-shaped-like-the-wrong-hypothesis.

F8. The three items marked "pending" in the blog skeletons all landed. N x N gave 134/134; the matched-null scree gave 11 above shuffled and 0 above permuted; stage 2 at seed 1 gave 15/15 and r = 0.978. Current: blog-skeletons, The N x N data-to-adapter scoring test, PCA, the scree curve and the two nulls, Stage two at a second seed.

F9. "Exactly predicted" holds only at stage 1. Stated: the r/d = 0.025 attenuation prediction was matched at 0.0265. Qualified by: stage 2 gives a slope of 0.116, 4.6 times the prediction, with a shared component every stage-2 adapter carries (981fa3b5:13044, 12919; the 2026-09-05 verdict addendum calls it unresolved). Current: stage-one-versus-stage-two-clarification, Stage two at a second seed; open in Open questions.

F10. "yep train all 40 please" - approved, never executed. Stated: on 2026-08-29 the assistant offered to train the six Lexicon words with no stage-1 adapter, taking the validation set from 34 to 40 (981fa3b5:3339, 10:28:06Z); Samuel agreed (981fa3b5:3356, 10:32:44Z). What happened: the zoo has 34 lexicon traits. qwen35/traits_secondary.json holds 40 words and constitutions.json a constitution for each; the six with no adapter are unconformable, frightened, busy, defenseless, significant, sleepy. The 2026-09-05 provenance answer (981fa3b5:12639) explains the gap as the constitution writer refusing six words as states rather than dispositions. Whether the refusal preceded or followed the approval is not recorded. Current: decisions-log, provenance-feedback, The six refused traits, The 34 Lexicon traits and how they were drawn; open in Open questions.

F11. The Imagination factor's Tucker congruence with Intellect, "about 0.6". Stated: 981fa3b5:13013. Replaced by the file: 0.6823, qwen35/results/fa_qwen35.json#solutions.centred_k5.congruence_oblimin, row 4 column 4 (target label order [E, A, C, ES, I, Eval]). Every other congruence in that answer matches the file to two decimals. Current: FA_Imagination (Imagination), Factor analysis of the adapter Gram.

8. Pre-zoo history

All from the Qwen2.5-3B era: the drift model organism, the gradient probe and the 100-trait sweep. They are historical, but they are reversals the project made and the blog post may narrate them.

Reading-page predictions corrected by running the experiment (2026-09-10)

Numbers corrected by re-deriving them from the per-item source (2026-09-10)

Predictions corrected by training on the data (2026-09-10)

Explanations corrected by removing the confound (2026-09-10)

Related: Source contradictions, Open questions, Superseded geometry claims, method-lessons, Timeline, How to read and maintain this wiki.

Display names changed while the result did not (2026-09-11)

The 2026-09-08 half of that row was not filed here when it happened; it is recorded now, with the 2026-09-11 rename, so the sequence of three display names is in one place.

Reframings between the 2026-09-10 draft and the 2026-09-15 short rewrite

Recorded on 2026-09-16 from qwen35/POST_DRAFT.md, its backups POST_DRAFT_2026-09-10_long.bak.md, POST_DRAFT_2026-09-15_long_illustrated.bak.md and POST_DRAFT_2026-09-15_preappendix.bak.md, and wiki/log.md. None is a reversal of a measured number; each is a change in what is presented as primary.

Reframings between the 2026-09-15 short draft and the 2026-10-30 post

Recorded on 2026-10-30 from wiki/raw/post-persona-geology-2026-10-30.md against qwen35/POST_DRAFT.md (post-draft). The post is now the current statement (Post: Persona Geology (2026-10-30)). No measured number was reversed; the differences between the post's figures and the files are in Source contradictions S30 to S35.