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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Open questions

What the project has not established - unexplained results, experiments never run, provenance that cannot be recovered, and decisions still owed by Samuel - each marked by kind and pointing at the page that documents it.

What the project has not established. Each entry is marked by kind:

• unexplained result - it was measured, it is real, and nothing accounts for it;
• never run - the experiment or analysis was specified and does not exist;
• provenance gap - a number or artefact exists but its origin cannot be recovered from anything on disk;
• decision pending - it is waiting on Samuel, not on data.

Companion pages: Superseded claims for claims the project reversed, Source contradictions for files that disagree.

In flight

Nothing is in flight as of 2026-09-16. The one item that was is closed below.

The full OCT persona replication. RESOLVED 2026-09-08. Was running as of 2026-09-07. Samuel asked for the stage-1 weight-space geometry to be replicated on the full OCT persona (stage-two merged) adapters. Modal jobs zoo-grampersona (134 x 134 exact persona Gram) and zoo-personaseed1 (15 exact seed-1 personas, then persona seed 0 against seed 1, seed-1 within, and stage-1 against persona); analysis in qwen35/analyse_fulloct.py. Results are on Full OCT persona replication: the persona arrangement is the stage-one arrangement (off-diagonal r 0.9915) and 15 of 15 second-seed personas identify themselves. Every geometry result on this wiki is still a stage-1 result unless it says otherwise - see stage-one-versus-stage-two-clarification.

Side effect worth knowing: the container meter had been static at $2,240.52 since 2026-09-05 and started climbing again when these jobs launched (qwen35/phase10_runs/zoo40_meter.log). Any spend figure quoted from now on needs its read time. See zoo-spend-ledger.

Unexplained results

The stage-two attenuation slope is 4.6 times the prediction. Stage one matches the analytic prediction: slope 0.0265 against r/d = 64/2560 = 0.025. Stage two at a second seed gives 0.116, with a shared component that every stage-two adapter carries. qwen35/PHASE3_VERDICT.md's 2026-09-05 addendum calls it unresolved and nothing since has explained it. Page: Stage two at a second seed; also The seed floor.

Stage-two separation is negative, and no document says so. qwen35/analysis/crossseed_arms_stage2.json#[0]: min_same = 0.04248138800846624 sits below max_off = 0.05140579106757132, giving sep = -0.00892440305910508. Stage one has positive separation in both arms, and the 2026-08-22 addendum makes "perfect separation, no overlap at all" one of its three structural facts. The 2026-09-05 stage-2 addendum reports 15/15 top-1 and does not report the separation at all. Page: Stage two at a second seed.

The matched shuffled null has a nominally significant clustering ARI. qwen35/results/decomposition_shuffled_matched.json#test2b."leading-component removed": ARI = 0.05977870383675816 at ari_p = 0.0009995002498750624. The original shuffled arm gave ARI = 0.00438485865621544 at ari_p = 0.36231884057971014. The 2026-09-05 addendum says the objective mismatch "changed nothing to three decimal places", which is true of the headline TEST 1B and TEST 2 and not of this. A structureless control should not cluster. Page: The null control arms.

Whether the hole is a real gap in the language or in the sample. The widest gap in the 134 words' coverage is 52.5 degrees at k = 5 (68.9 in the full sketch space). Three candidate English words were trained for it and do not agree with each other; the activation-space transplant works against a permuted null. The page's verdict is "suggestively, not convincingly, insouciant", and the project's own reading is now "a direction the adapters leave open" rather than "a character English has no word for". Status on the page is unconfirmed. Page: Where no word goes - the hole and the alien direction; the withdrawal is Superseded claims item F5.

Extraversion and Emotional Stability come out rotated, and nobody knows how stable that is. Three of five Big Five axes recover cleanly; E and ES emerge as withdrawal and arousal instead. No Goldberg congruence clears 0.85 in any solution. Whether the rotation is a property of the adapter cloud, of the oblimin rotation, or of the 6/14 keying imbalance in Emotional Stability (see Source contradictions S10) has not been tested. Pages: explainer-big-five-mapping, Factor analysis of the adapter Gram, axis_Extraversion (the named Extraversion axis), axis_EmotionalStability (the named Emotional Stability axis).

The STEER134 coherence-collapse argument. qwen35/analysis/manifold_ideas.md builds an argument about coherence collapsing at high alpha from qwen35/results/steer134_judged.json, whose coverage.complete is false: 11,010 of 28,664 units are unjudged and many cells sit at n = 4-7 against n_expected 12, against only two random controls. Marked unconfirmed. Either finish the judging or drop the argument. Page: Steering the base model along weight-space directions.

LoRA-A drifts 4.5% over 93 steps, somewhere. qwen35/build_monitor_page.py line 112 says so; every analysis JSON says 0.0146 / 1.5%. Whether 4.5% is a different population or a superseded measurement is not recoverable. Source contradictions S4; page Adapter effect and LoRA-A drift.

Is Persona Cartography's PC10 the same object as the Assistant Axis? PC10 "cleanly separates the baseline (the model with no LoRAs) from the others" and reads qualitatively as self-reference. A base-versus-personas direction is what a grand-mean direction is, up to sign and centring, and qwen35/steer_qwen35.py calls the grand mean the weight-space analogue of the Assistant Axis. One qualitative paragraph on 11 points against an activation-space result is not an identity. Written up as a resemblance. Page: The Assistant Axis (Lu et al., 2026).

Never run

The enumerated-anchor ablation. constitution_enumerated exists on every accepted entry of qwen35/constitutions.json and qwen35/anchor_constitutions.py describes the arm, but no corresponding corpus and no phase2_runs/results_* file exists. This is the ablation that would show whether naming one Goldberg marker per factor in the anchor could have manufactured the Big Five result. Page: The anchor block and its 2026-08-19 revision.

The neutral-constitution null. qwen35/paper_notes.md section 3.9 recommends it as "the null that a reviewer will ask for". The phrase appears nowhere else in the repository. Page: Where the recipe diverges from the two source papers, The null control arms.

The teacher screen was paid for and never analysed. teacherscreen/results/generations.jsonl (2,880 rows, $0.119453) and judgements.jsonl (5,760 rows, $8.820759) are on disk; teacherscreen/analyze.py writes results/teacher_screen.{json,md} and neither file exists. So the screen's own questions - where a teacher starts failing by tier, whether amplify and suppress separate equally, whether the six-word synonym cluster is one direction - are unanswered on this project's own paid-for data. One script run away, but running it produces new numbers. Page: The teacher screen.

RESOLVED 2026-09-08 - the reward-hacks arms now have a behavioural comparison. The entry read: the control arm produced no file, sorh_eval.log ended mid-run at [checkpoint-31] done, no judged_sorh*.json existed, and any claim about reward hacking's behavioural effect was unsupported. The cause was a manual systemctl stop at 2026-09-01T14:34:24, not a missing adapter. zoo-sorheval2.service re-ran the control arm on 2026-09-08 (qwen35/phase10_runs/rl_persona_sorh_control.json), both arms were judged blind (qwen35/phase10_runs/judged_sorh.json, 192 records, 0 failed calls) and analysed (qwen35/analysis/sorh_behavioural.json). The answer is a null: both arms fall about a point of judged Conscientiousness and Intellect below base, but nothing separates hack from control after Holm correction. Page: The reward-hacks arms — the positive control that also failed.

The reward-hacks behavioural effect is confounded with response length. Both arms collapse from 305.4167 words to about 70 (qwen35/analysis/sorh_behavioural.json#text), and the judge's Conscientiousness and Intellect scores correlate with cleaned response length at Spearman 0.4942 and 0.3965 over the 192 records (#length_vs_factor_spearman.rho) - exactly the two factors that dropped. A length-matched decode would be needed to say the fine-tune changed personality rather than verbosity. unexplained result. Page: The reward-hacks arms — the positive control that also failed.

The math+code GRPO run stopped at step 99 of 500. qwen35/phase10_runs/rl_mix.log ends there and no analysis/rl_sketches_mathcode.json exists. It is also the only RL arm that adopted the zoo's LoRA-A (adopted zoo LoRA-A on 248/248), so it is the only one whose weight-space comparison to the zoo is on a shared basis. Page: Capability RL and persona drift.

A Frobenius-norm table comparable to Persona Cartography's Table 1. qwen35/paper_notes.md section 3.8 calls it "near-free and PC reports it", and says that if this project's 140 norms are not in a tight band then every "sum the scales" intuition from that paper is invalid here. No such table exists in qwen35/analysis/*.json. Now that rsLoRA is off and the scaling conventions match (Superseded claims item L-1), the comparison to their 6.08-6.53 band is directly meaningful. The cheapest remaining cross-paper number. Pages: Persona Cartography (Baines et al., 2026), Where the recipe diverges from the two source papers.

axis_Conscientiousness's preregistered prediction was never scored. The registered criterion was "negative arm effect larger in magnitude than positive arm"; no file computes an arm-asymmetry statistic. Page: Steering the base model along weight-space directions.

Stage-two adapters were never run in activation space, and the activation arm produced no behavioural measurement at all - four responses per trait were kept for eyeballing and none was judged. Pages: Activation-space analyses: overview, Activation-space analyses: method notes and limits.

No weight-space cross-seed Procrustes fit. The only Procrustes in the repository is in activation space (qwen35/analysis/actspace_geometry.json#windows.resp.primary.procrustes_r2). The cross-seed arrangement result rests entirely on the Gram correlation and the replicated decomposition. Page: Cross-seed geometry.

An 8B replication arm (paper_notes.md 3.3) and a LIMA general-prompt arm (3.5). Neither exists; 3.5 is the largest unresolved divergence from both source papers. Page: Where the recipe diverges from the two source papers.

The experimental plan's largest items. personality_lora_zoo_experimental_plan.md (repository root) Tier 1 items 7, 8 and 9 - PLS / reduced-rank regression, cross-validated performance against subspace dimension k, representation ablation - were not done, and H6, the causal test (project onto the subspace and project it out), was never attempted. That is the largest single gap against the plan the project set itself. Page: The experimental plan, and what was done against it.

All four of CONTEXT.md's open questions. Can retrievable content be made to speak; does DPO produce something prompting does not; does the KL result survive a non-saturated task; a second seed for the saturation decay curve. None was run. Pages: Origin, and how the question changed shape three times, The drift experiment — weight-space control of a trait fails, The gradient-content probe — reading works, but only during acquisition.

The DPO-epochs re-check at LR 5e-06. "More DPO epochs does nothing" stands as measured, but the learning rate was 5e-05 throughout where guidance is about 5e-06, and an LR ten times high produces the same saturation signature. The cheap re-check was never run. Page: The 100-trait sweep on Qwen2.5-3B.

The N x N heatmap figure. (Still not built as of 2026-09-16; the companion's methods page prints the 134 of 134 statistic and the post carries it as Appendix A8.) Named in chat as the one genuinely new figure the site needs (981fa3b5:13021, 13065). The matrix exists in qwen35/analysis/nxn_summary.json; no figure has been built. Page: The N x N data-to-adapter scoring test.

A literature search for prior art doing factor analysis on weight deltas. paper_notes.md section 4 item 8 records it as not found, and it was not a search. Still not done. Page: Factor-analysis methods reference.

Provenance gaps

Eight load-bearing JSON files have no producing script anywhere in the repository. qwen35/analysis/module_holography.json, polarity_deflation.json, adapter_effect.json, intrinsic_coords.json, steerfix_replication.json, judged_ceiling.json, functional_probe.json, qual_pairs.json. Also read but never written by checked-in code: page_data.json and monitor_page_data.json. Consequences by file:

• steerfix_replication.json carries the entire steering replication table and is read by two builders; the definitions of slope2, sel2, coh2 and named are reconstructed from consumers, not stated anywhere. Steering the base model along weight-space directions.
• module_holography.json is the only source for the one-module-classifies-factor result (0.638 against chance 0.202, all 248 reaching 0.76). Module holography.
• polarity_deflation.json is where 4.6% comes from and the 46% figure cannot be derived from it. Polarity, bipolarity and the trait graph; Source contradictions S18.
• judged_ceiling.json (observed 0.582, frac_of_ceiling 0.704) and functional_probe.json (0.581 within / 0.129 across, retrieval 0.032) carry the judge-reliability ceiling. sb_full at least checks out arithmetically as the Spearman-Brown extrapolation of split_half_mean. Judged evaluations of the trait adapters.
• adapter_effect.json is owned by nobody. The geometry section concluded it is probably behavioural and passed it to behaviour; behaviour cited it only as a file that exists; the trait generator printed its five fields verbatim with a statement that their meaning is not established. Its keys (sim_base, sim_s1, rep, leak, chars) are undefined and nothing on the blog page reads it. Adapter effect and LoRA-A drift, Judged evaluations of the trait adapters.
• intrinsic_coords.json has neither producer nor consumer. PCA, the scree curve and the two nulls.

The sketch-validation correlation is stored nowhere. Two built pages quote two different numbers for it (0.9996 and 0.99944) and no file holds either. It is the justification for reading sketch angles as weight-update angles. Source contradictions S5; page Weight-space geometry of the 134-adapter zoo.

The scree real curve was never regenerated from the exact Gram. #real is byte-for-byte identical in analysis/scree_null.json and analysis/scree_null_matched.json, and equals analysis/geometry_stage1.json#explained_var_top10, which is computed over the 100 labelled traits from the stage1_k32 sketches. The finding is recorded in .garden/journal/2026-09-05.md and nowhere in the data. Either regenerate it from results/gram_sweep.npz or write one line into the JSON saying it is deliberately the 100-trait sketch spectrum. Page: PCA, the scree curve and the two nulls.

The per-alpha damage counts for the five axis_* directions. analysis/qual_axes.json has no damage_markers.per_alpha.looping where the other three qual files do, so blog_data.json#degen is empty for those five and the blog's "intact over alpha -2, -1, 1, 2" prints from an absent record. qual_axes_notes.md says the counts were made programmatically across all 24 prompts at all 7 alphas. They are not in the JSON. Page: Steering the base model along weight-space directions.

Condition-level judged means for base, stage 1 and persona. qwen35/phase10_runs/judged_100.json holds all 7,200 individual judge records; no file stores the three condition means. analysis/spider.json#base_trait is the only sourced condition mean found. Wiki rule 1 forbids recomputing them, so no persona-minus-base effect size is quoted anywhere. Page: Judged evaluations of the trait adapters.

The activation cross analysis printed its class medians and never saved them. qwen35/analyse_actspace_cross.py prints full per-class median tables; no log in phase10_runs/ has them. The medians survive only in qwen35/ACTSPACE_RESULTS.md and the rendered blog page - and the same factor, same keying class (n = 14) has no published medians at all. Page: Adapter x constitution, 16 x 16.

results/runmeta_actspace.json does not exist, so decompose TEST 6 (the training-strength nuisance covariate) is unavailable for the activation Gram. The log says it plainly: "The confound is then UNTESTED, which is not the same as ruled out." Page: Activation-space analyses: method notes and limits.

The seven later adapters were trained on unanchored constitutions, against a document that says "same recipe". qwen35/PREREG_alignment.md says the four alignment adapters use "Same recipe as the 134-adapter zoo". Their constitutions, and the three hole ones, were generated after the 2026-08-19 anchoring migration and never put through it: their constitutions.json entries carry only a constitution field and the text does not end with the generic anchor paragraph that all 134 carry. gen_pairs.py conditions the teacher on that field. Page: The alignment and hole traits.

Whether the Hugging Face model repository is public. qwen35/upload_zoo_batched.py:177 prints "Repo still PRIVATE." as a hardcoded literal, not a live check. The dataset repo's status is file-sourced (analysis/hf_dataset_audit.json#private: false). No file records a model-repo inventory or visibility read after 2026-08-29. Nor does anything record whether the alignment and hole adapters are on the Hub at all - upload_zoo_batched.py builds its job list from the pc-qwen35-sweep and pc-qwen35-oct2 volumes only, which do not contain them. Update 2026-09-16: the alignment and hole adapters are now on the Hub in the controls repo, whose visibility was read live (private=False) rather than printed as a literal; the zoo repo's own visibility is still not file-sourced. Pages: Hugging Face artefacts, The alignment and hole traits.

Two cost files were overwritten by later runs. qwen35/constitutions_cost.json was overwritten by the three-trait hole run ($0.013221, 3 calls), so the 140-trait constitution cost exists nowhere. The zoo's own data/_usage.json and _dropped.json were overwritten by the alignment run; the zoo's pair-generation totals survive only in genpairs*.log, which give four per-pass figures and no total. Pages: Constitution generation, DPO pair generation, zoo-spend-ledger.

One week has no narrative record. 2026-08-25 to 2026-08-31: the harness was stopped 2026-08-25 11:44, v3 went live 2026-08-31, and the project journal begins 2026-09-01. The batch 2/3/4 stage-two runs, the content review, the thinking-default discovery, the HF rate limit and the budget raise are dated only by file mtimes and two auto-memory notes. Page: Timeline, harness-context.

Which stage-two batch warm went through. It is in no zoo-*.service unit; it appears in the three-trait pilot results file and in merge_audit.json, and is inferred from those two. Page: Stage two - introspection SFT and the 0.25 persona merge.

Why the six drawn-but-unbuilt lexicon words were dropped. No source records it. Source contradictions S11; pages The six refused traits, The 34 Lexicon traits and how they were drawn. Resolved 2026-09-10: The six refused traits quotes the writer's NOT_A_TRAIT reason for each of the six from the pre-anchor backup; what stays open is F10 on Superseded claims.

Where the base model came from. The choice of Qwen3.5-4B is not made in any of the thirteen Claude Code transcripts; the earliest project turn already asks what the adapters were trained on. It predates 2026-08-22 and would have to be recovered from the repository or the old harness memory. Likewise the choice of Goldberg's 100 unipolar markers and the draw of the 40 lexicon words: Samuel asks about the set twice as a question, never as a decision. Page: decisions-log, Origin, and how the question changed shape three times.

Whether Emotional Stability's 6/14 keying is Goldberg's or a transcription error. Not checked, and it affects every per-factor statistic that assumes balanced keying. Source contradictions S10; page The 100 Goldberg primary traits.

The ordinal labels in the "objective must travel" lesson are inferred. The sources give counts - the 2026-08-24 PHASE3_VERDICT.md addendum says "the third occurrence of this project's signature failure mode" and .garden/journal/2026-09-04.md says "the fourth" - but nothing enumerates which two came first. SW_BASE_MODEL and PC_LORA_ALPHA were assigned to slots 1 and 2 on chronology alone; PC_USE_RSLORA is an equally plausible earlier member. All four incidents are real and sourced; only the ordering is inferred. Page: lesson-objective-must-travel.

Persona Cartography's venue is unconfirmed (the checkout carries paper/neurips_2026.sty and a "camera-ready switch" HEAD commit, and no acceptance record was found), and on Big Five reference bibliography eleven entries are marked [not verified] against six [verified] - the eleven were written from general reference, not confirmed against a publisher record. Pages: Persona Cartography (Baines et al., 2026), Big Five reference bibliography.

Several pre-zoo numbers survive only in the old harness memory. results/pca_over_loras.txt and results_pca.log are zero bytes, so the PCA-over-early-LoRAs spectra, the rank-spectrum energies, the interpolability figures, the rank-truncation series, the A-GEM constraint diagnostics and the saturation quartile decay exist only in ~/projects/agent-harness/memory/projects/persona-curvature.md and the repository-root CONTEXT.md. Pages: Early LoRA structure — the 34 OCEAN adapters, The drift experiment — weight-space control of a trait fails, The gradient-content probe — reading works, but only during acquisition, costs.

Decisions pending

All are Samuel's, and none is waiting on data.

Co-authorship for the external reviewer. Raised in the review and not answered. Page: external-review, decisions-log.

The Persona Cartography reframing the review asked for. Partial credit exists on the live page; the fuller rewrite does not. Page: external-review.

Hosting, and a static LessWrong version. The review asked for a hosted version. As of 2026-09-07 the zoo's blog page is not served: the three Caddy routes in /etc/caddy/Caddyfile are this wiki, the project server on port 8091 serving the historical sweep100/site2/, and the project server on 8092 serving qwen35/site/, the phase plan page. Page: external-review, blog-site-map-idea, infrastructure.

The short-post-plus-website plan, and per-trait pages. The ten-page site map of 2026-09-07 is a proposal; the site has not been built. The 141 trait pages in this wiki are the closest thing that exists. Page: blog-site-map-idea, Trait index.

"yep train all 40 please" - approved 2026-08-29, never executed. The zoo has 34 lexicon traits. Whether to train the remaining six, or to close the instruction as overtaken by the constitution writer's refusals, is undecided. Pages: decisions-log, provenance-feedback, The six refused traits; history at Superseded claims item F10.

The content-publication policy. qwen35/PENDING-CONTENT-REVIEW.md puts three options to Samuel and recommends option 2 (publish minus the quarantined files), specifically over self-harm-pattern content in the stage-two corpus (self_interaction/temperamental.jsonl at 10.4%). The 2026-08-30 upload log shows option 1 in effect - everything uploaded, previously flagged rows cleared - and qwen35/phase10_runs/adjudications.json holds 60 written clearance decisions. Nothing on disk records a decision by Samuel. This is the only open decision that has already been acted on. Pages: zoo-build-governance, Hugging Face artefacts.

Whether to re-cut the two published Hugging Face cards. Both are stale and public; see Superseded claims section 1d. Page: Hugging Face artefacts.

Related: Superseded claims, Source contradictions, method-lessons, The experimental plan, and what was done against it, How to read and maintain this wiki.

Added 2026-09-08

• Resolved: published exact personas were inert. The persona_exact/ adapters on the Hub carried a doubled tensor-key prefix and loaded through PEFT with no effect (tested). Repaired on the volume and re-uploaded with Samuel's approval; verified by header reads. See Full OCT persona replication, Hugging Face artefacts, lesson-verify-the-artefact-loads.
• Unexplained result: the stage-two shared direction. Fifteen percent of every stage-two adapter's squared norm lies along one common direction, orthogonal to anything in stage one. No behavioural test and no control SFT exists for it. See Structure of the stage-two adapter space.
• Never run: stage-two factoring with the shared direction projected out, and a control SFT on self-generated but trait-free transcripts. See Structure of the stage-two adapter space.
• Never run: factor analysis at the second LoRA seed. Whether independently factoring the 40 seed-1 adapters recovers the same five factors. The per-direction check that was run instead is uninformative; the matched-objective arm has no 40 x 40 within-arm Gram on disk. See Per-direction seed stability, and why the check is uninformative.
• Run 2026-09-12: a 100-trait real arm for the null comparison. The two null arms have 100 variables against the real arm's 134, so their eigenvalues were not on the same scale; factoring the 100 Goldberg markers alone out of results/gram_sweep.npz fixes it. Done in Goldberg-only factoring and the 34 as a held-out set: the p = 100 real arm's centred spectrum is 12.725, 11.140, 5.663, 3.996, 2.927 against the permuted arm's 13.632, 10.390, 4.945, 3.887, 2.824, and both retain 8 factors, so at equal p the spectrum does not separate real labels from permuted ones and only congruence does. See Factor analysis of the null arms.
• Never run: additivity for PC and factor mixtures. Mixture additivity exists only for the five Big Five keying axes. See What has been done to the PCs but not to the factors.
• Never run: a sphere sweep sampled in the factor subspace. The 72 judged directions were sampled on the top-three-PC sphere; only the layout has been moved to the factor chart. See What has been done to the PCs but not to the factors.

Reading, 2026-09-09

• Two papers read in full and related to the project, with eleven ranked experiments they suggest: SliderSpace (Gandikota et al., ICCV 2025) and Gradient Atoms (Rosser, arXiv:2603.14665, 2026). The cheapest and most consequential is redoing the factor analysis in the Fisher metric rather than the Frobenius one, which both papers' methods imply and which Fisher norms of the steering directions has already measured the need for. See Paper reading, 2026-09-09 - SliderSpace and Gradient Atoms.
• Run since (2026-09-10): G2, G4 and G3. Gradient atoms on the zoo's own preference data and Gradient atoms on the reward-hacks corpus. Still never run from that page: G1 (the factor analysis in the Fisher metric, the item above), G5, G6, S2, S3, S4 and S5. Two questions the run opened rather than closed: no atom was ever steered, so the paper's second half is untested here; and the whitening used is the empirical Fisher on realised tokens over document gradients, so whether a different approximation gives a different dictionary is unknown.

Added 2026-09-16, from the freshness audit against the 2026-09-15 draft

• Never run, named by the draft as the next three runs ("What we would run next"): constitutions written without Big Five vocabulary, to test whether the rotation and the five-factor count are the teacher's; the same 134 datasets on a second base model; the enumerated-anchor ablation (The anchor block and its 2026-08-19 revision). The fourth item on that list, the self-identification probe, was run on 2026-09-15 (Self-identification probe).
• Unresolved divergence from the source papers: 13 optimizer steps against OCT's roughly 47. Recorded on Where the recipe diverges from the two source papers section 3.5 and pointed to from Stage-one training configuration; the draft states it in its scope note and does not resolve it. The cheap test (one trait trained with and without a LIMA-style general pool) has not been run.
• Unexplained: parallel analysis retains nine factors, the stage-two residual seven, the permuted null eight. Five is the hypothesis. The draft says so ("That five is the number" under "What this does not show"); no analysis decides between five and nine. Pages: Factor analysis of the adapter Gram, PCA, the scree curve and the two nulls.
• Unconfirmed: the reward-hack arms' behavioural null is length-confounded (both arms collapse to about 70 words; see the entry above). A length-matched decode is still the missing control. Page: The reward-hacks arms — the positive control that also failed.
• Provenance: two analysis files the draft's figures depend on have no producing script checked in under qwen35/. analysis/best_axis_pairs.json (Figure 1) and analysis/fa_text_contrast.json (the text-contrast factors) were written by session scripts on 2026-09-15; on 2026-09-16 the scripts were checked in as qwen35/analyse_best_axis_pairs.py and qwen35/analyse_fa_text_contrast.py and shown to regenerate both files exactly. Resolved. See Code and data map for the remaining files without a producer.
• Resolved 2026-09-16: the non-weight data release. The analysis, results, corpora and run-log files are public at https://huggingface.co/datasets/EternalRecursion/persona-curvature-results (3,384 files, 3.94 GB) and the GitHub repository https://github.com/EternalRecursion121/persona-curvature carries tools/fetch_data.py and a sha256 manifest. The control adapters followed on 2026-09-16 with the owner's approval: EternalRecursion/persona-lora-zoo-qwen35-controls, 353 adapters, 162.57 GB, verified against qwen35/analysis/hf_controls_manifest.json. Still on request: the unmatched first null run, the seed-1 stage-two adapters, the four pilot traits and the phase-2 bake-off runs. Page: Code and data map, Hugging Face artefacts.
• Decision pending: the wiki and companion are public without basic auth (removed at Samuel's request on 2026-09-07; confirmed in /etc/caddy/Caddyfile on 2026-09-16). The wiki carries spend figures, transcript-sourced pages and an unpublished review's paraphrase. Page: How to read and maintain this wiki.
• Numbers the 2026-09-16 audit could not verify against a file. Listed in pages/overview/_report_freshness_2026-09-16.md; the load-bearing ones are the sketch-validation correlation (0.9996 or 0.99944, provenance gap above), the hard-coded blog-page strings in Superseded claims section 1c, and the literature numbers on the paper pages, which cite arXiv URLs rather than files on disk.