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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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.
Claims and evidence
The seven-row claims table of the 2026-09-15 post draft expanded so that every number in it names the analysis file and JSON key it was read from on 2026-09-16, the wiki page that holds the full result, and the control that supports it.
currentverified 2026-09-16
- Row 1. The updates have five-factor structure
- Row 2. The structure is the training contrast's
- Row 3. A trait is its output subspace
- Row 4. Map is not behaviour
- Row 5. Stage two installs a register first
- Row 6. Direction predicts personality
- Row 7. First-order score flags harmful data
- What the draft states as method, and where it is recorded
- What this page does not do
Since 2026-10-30 the current statement is the post Persona Geology (Post: Persona Geology (2026-10-30)), which has no claims table; its numbers were checked against these rows and the differences are in Source contradictions S30 to S35. The post draft of 2026-09-15 (post-draft) closes with a seven-row table: claim, evidence, status, what would break it. This page is that table with the evidence column opened up. Every number below was re-read from the named file and key on 2026-09-16; where the draft rounds, the full-precision value is given beside it. The "control" column names the measurement that rules out the obvious alternative explanation. Where the draft's number is a ratio the file does not store, both operands are cited and the arithmetic is shown.
Conventions: results/ and analysis/ are under qwen35/. "Tucker congruence"
and the other terms are defined in the Glossary. The wiki page named in each
row holds the full result, its caveats and its history.
Row 1. The updates have five-factor structure
Status in the draft: holds; Big Five congruence 0.40 to 0.68. Would break it: independent constitutions giving a different rotation.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| Warmth to Agreeableness 0.66 | 0.6554774644183582 | results/fa_qwen35.json#solutions.centred_k5.congruence_oblimin[0][1] |
Factor analysis of the adapter Gram, FA_Warmth (Warmth / prosociality) | permuted-label arm: agreement with labels at chance |
| Competence to Conscientiousness 0.57 | 0.5744454556114018 | ...congruence_oblimin[1][2] |
FA_Competence (Competence) | same |
| Timidity to Emotional Stability 0.40 | 0.4048918307420467 (and 0.3380 to Extraversion) | ...congruence_oblimin[2][3], [2][0] |
FA_FearfulWithdrawal (Timidity) | the rotation is also in the text (row 2) |
| Arousal to Extraversion 0.54 | 0.538926935263515 (and -0.3483 to Emotional Stability) | ...congruence_oblimin[3][0], [3][3] |
FA_Arousal (Arousal / activation) | same |
| Imagination to Intellect 0.68 | 0.6822835176928186 | ...congruence_oblimin[4][4] |
FA_Imagination (Imagination) | same |
| parallel analysis retains nine | chosen 9 at reference N 1528 |
results/fa_qwen35.json#n_factors.chosen, #n_factors.parallel_analysis_centred.grid[3].k_unreduced_95pct |
Factor analysis of the adapter Gram, PCA, the scree curve and the two nulls | five is the hypothesis, not the data's count |
| same-pole +0.24, opposite-pole -0.08 | 0.24472701836565558, -0.08134207180636474 | analysis/stage2_structure.json#decomposition_tests.stage1.test2_same_opp_gap_p[0..1] |
Polarity, bipolarity and the trait graph | permutation p 4.999750012499375e-05 (same key, index 3) |
| shuffled arm retains zero factors | 0 | analysis/fa_nulls.json#arms.shuffled.n_factors_chosen |
Factor analysis of the null arms, The null control arms | trained at the zoo's exact objective on the same 445 prompts |
| permuted arm retains eight, agreement at chance | 8; best-matched congruence mean 0.1779 | analysis/fa_nulls.json#arms.permuted.n_factors_chosen, #tucker_vs_real_stage_one.permuted.mean_abs_best_matched_congruence |
Factor analysis of the null arms | re-identified by data (next row) |
| re-identified permuted factors at 0.98 to 1.00 | 0.9976, 0.9951, 0.9854, 0.9940, 0.9838 (signs free) | analysis/fa_nulls.json#tucker_vs_real_stage_one.permuted.source_relabelled.best_matching[*].congruence |
Factor analysis of the null arms | the shared prompt pool is not what the factors are made of |
| Fisher metric 0.97 expected, 0.96 empirical | 0.965359016717984 and 0.9577766396647475 (mean abs matched) | analysis/fa_fisher_metric.json#loading_matrix_congruence_k5.pairs.frobenius_vs_fisher_expected.mean_abs_matched_congruence, ...frobenius_vs_fisher_empirical.mean_abs_matched_congruence |
The factor analysis in the model's own metric | metric change, not a rerun |
| Goldberg-only factoring 0.99 | 0.9922779568528804 | analysis/goldberg_only.json#verdicts.T1a_factors_survive_dropping_the_34.value |
Goldberg-only factoring and the 34 as a held-out set | the 34 held-out words are placed where an independent rater puts them (same page, test 2b) |
Row 2. The structure is the training contrast's
Status: holds. Would break it: a second base model landing traits elsewhere on the same data.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| arrangement complete at rank 1 | chart Pearson 0.9967445664202496; 15 of 15 identified | analysis/rank_sweep.json#ranks.r1.chart.pearson_75, #ranks.r1.identification_among_134.top1 |
Rank sweep | rank-1 judged own-factor movement -0.41 against 27.08 at rank 64 (#behaviour.ranks.r1.mean_own_factor_amplification, ..._rank64) |
| text readout predicts 94% of held-out position | held-out R^2 0.9380347105286267 | analysis/goldberg_only.json#test3_text_embedding_baseline.ridge_pair_contrast_to_chart.heldout_r2_overall |
Goldberg-only factoring and the 34 as a held-out set | the constitution alone: held-out R^2 -0.034409160222032664 (...ridge_constitution_to_chart.heldout_r2_overall) |
| adapter geometry itself 98% | 0.9822421695022922 | ...comparator_adapter_geometry.r2_overall |
same | ceiling for the text readout |
| text Gram and adapter Gram correlate 0.86 | 0.8601935162272526 | ...text_gram_vs_adapter_gram.max_of_constitution_and_contrast |
same | |
| text factors match adapter factors 0.81 to 0.97 | Warmth 0.9230, Imagination 0.9023, Competence 0.9705, Timidity 0.8131, Arousal 0.8967; mean 0.901112105638296 | analysis/fa_text_contrast.json#best_matching, #mean_abs_matched_congruence |
Factor analysis of the training-contrast embeddings | the matching is a permutation; the Timidity and Arousal rotation appears in the text solution too |
| prompted arrangement, Procrustes 74%, null 4% | 0.7387431438763885; null 95th percentile 0.03568178186070537 | analysis/actspace_geometry_fa.json#windows.resp.procrustes_r2_fa, #windows.resp.procrustes_null_95pct |
Constitution-as-prompt persona vectors vs weight geometry | an artefact check, not evidence the model added structure (draft A7) |
Row 3. A trait is its output subspace
Status: holds within one recipe. Would break it: cross-seed column-space overlap failing on new data.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| cross-seed cosine 0.018 | 0.018060998368198355 (n 40) | analysis/act_gram.json#arms.frob.a_same_trait_cross_seed_mean |
The seed floor | r/d = 64/2560 = 0.025; frame overlap 0.021471366484559353 (#arms.frob.c2_frame_overlap.seed0_vs_seed1.c2b_summed) |
| 40 of 40 identified; matrices agree 0.997 | 40; 0.9966259139824906 | #arms.frob.d_top1_of_134, #arms.frob.e_pearson_cross_vs_within |
Cross-seed geometry | different-trait cross-seed cosine 0.0018594300142914396 (#arms.frob.b_diff_trait_cross_seed_mean) |
| activation-weighted cosine 0.66, unrelated 0.09 | 0.6642887281525424; 0.08760415842065107 | #arms.pool445.a_same_trait_cross_seed_mean, #arms.pool445.within_seed0_offdiag_mean_cosine |
The activation-weighted Gram, and what the seed floor looks like in it | ceiling 0.8909847593616453 (#verdict.null_c2b_C); pairs at the same fraction of ceiling in both metrics (#verdict.ratio_a_C_over_null_C 0.7456, #verdict.ratio_a_F_over_null_F 0.8412) |
| column space 46% vs 14% | same_mean 0.46279728915356094, off_mean 0.14087331941381756, top1 40 of 40 | analysis/column_space.json#stage1.identification_40_seed1_queries_vs_134_seed0.col_wtd_k64 |
Column-space structure | row space identifies 19 of 40 (...row_unw_k64.top1) |
| top output directions agree 0.77 vs 0.21 | 0.7701377220108219; 0.21012838847727913 | #stage1.top1_left_vector_same_trait_cross_seed.mean_abs_cos, #stage1.top1_left_vector_diff_trait_cross_seed.mean_abs_cos |
Column-space structure |
Row 4. Map is not behaviour
Status: holds. Would break it: a metric in which functionally similar adapters are always close.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| rank sweep | see row 2 | analysis/rank_sweep.json |
Rank sweep | |
| sliders at cosine 0.003 to 0.022 with the trained adapter | 0.0025055221550983157 (warm) to 0.022155165412935125 (organized) | analysis/persona_sliders.json#weight_space.sliders.<trait>.cos_with_134.<trait> |
Persona sliders - LoRAs trained to an activation-space target | the sliders move the judged persona the right way (#behaviour.own_factor_movement) |
| stage one vs stage two of one trait, activation-weighted 0.009 | R_res 0.009222883147729255 | analysis/act_gram_stage2.json#verdict.R_res |
The stage boundary in the activation-weighted metric | same-function benchmark 0.7455668811085819 (#verdict.benchmark_same_function_stage1_cross_seed) |
| Fisher norms span a factor of 260 | 259.924881015629 | analysis/fisher_norms.json#comparisons.full_range.ratio |
Fisher norms of the steering directions | random-merge band median 0.12382989334918551 (#random_band.median) |
| Warmth three times a random merge | 0.36503759829531435 / 0.12382989334918551 = 2.95 | #directions.FA_Warmth.F_ref, #random_band.median |
Fisher norms of the steering directions | |
| stage-two shared direction 6.7 times flatter | 0.14899427326654663 (its inverse is 6.71) | #comparisons.stage2_vs_stage1_grand_mean.ratio_S2mean_over_stage1mean |
Fisher norms of the steering directions, What the stage-two shared direction does |
Row 5. Stage two installs a register first
Status: holds. Would break it: a persona whose extra amplitude comes from the register.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| 15% of every stage-two adapter's norm; cosine 0.39, spread 0.03 | 0.1514038882692439; 0.38929785188872645; sd 0.031056220786483055 | analysis/stage2_structure.json#shared_component.stage2 |
Structure of the stage-two adapter space | stage one: 0.07948767279497902, 0.28103697835548513, sd 0.12559260390449356 (#shared_component.stage1) |
| trait-free runs at 0.30 | 0.30353526531657493 (five seeds, sd 0.0007) | analysis/stage2_neutral_control.json#neutral_vs_grand_mean.mean |
Stage two, explored | nearest zoo trait unemotional for all five (#per_neutral[*].max_cos_trait) |
| residual correlates with stage one at 0.81 | 0.811787783969743 | analysis/stage2_structure.json#centred_cosines.corr_stage1_stage2 |
Structure of the stage-two adapter space | |
| bipolarity mostly gone, opposite poles +0.12 | 0.12301499153773737 (same-pole 0.19104548852552697) | #decomposition_tests.stage2.test2_same_opp_gap_p[1], [0] |
Polarity, bipolarity and the trait graph | stage one -0.08134207180636474 |
| persona arrangement is stage one's, 0.99 | 0.9915249945319294 | analysis/fulloct_geometry.json#gram_correlation_offdiag.persona_vs_stage1 |
Full OCT persona replication | |
| shared direction accounts for 0.9% of the extra amplitude | 0.009326363320103694 | analysis/stage2_exploration.json#register_vs_residual.share_of_gap_cond_c |
Stage two, explored | |
| stage one names its word 0.2%, persona 0.3%, stage two 13 traits and chart cosine 0.61 | 0.0023320895522388058; 0.002798507462686567; 13; 0.6105703585726646; same factor and pole 0.5584415584415584 | analysis/selfid.json#per_prompt.P1.conditions.{stage1,persona,stage2}.{exact_hit_rate,n_traits_with_any_exact_hit,chart_cos_mean,same_factor_and_pole_share} |
Self-identification probe | permutation null exact-hit p95 0.0032649253731343282 (...stage2.exact_hit_rate_null_p95); the pre-registered persona prediction failed |
Row 6. Direction predicts personality
Status: holds in the zoo's frame. Would break it: steering at matched dose that breaks the sphere correlation.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| sphere Spearman 0.68; 0.95 near, 2.92 far; 45 of 72 habitable | 0.6765886845997277; 0.9462620103993861; 2.917385198267297; 45 | analysis/sphere_page_fa.json#smooth.{rho,near_mean,far_mean}, #coherence.none |
The sphere sweep on the factor chart | iso-KL rerun of the PC sphere keeps 0.65 (The sphere sweep at iso-KL dose) |
| suppress-over-amplify 1.90 at equal alpha, 1.65 at matched dose, 0.80 by headroom | 1.8957528957528955; 1.6455696202531644; 0.7946215813581846 | analysis/matched_dose_steering.json#published_equal_alpha_summary.all.suppress_over_amplify_raw, #summary.all.suppress_over_amplify_raw, #summary.all.suppress_over_amplify_room |
Suppression versus amplification at matched Fisher dose | KL target 0.24765978669837102 nats per token (#kl_target_nats_per_token) |
| dials: 8 of 10 (axes), 10 of 10 right direction and 8 most (factor adapters) | counts on the pages | analysis/spider.json#axes, #bigfive |
OCEAN dials replication (spider plots), Big Five factor adapters (Persona Cartography's own ten dials) | keyed-adjective average 7 of 10, personas 8 of 10 (#traits, #personas) |
| additivity deviation 53% | 0.5282408112973758 | analysis/additivity.json#median_resid_over_effect |
Additivity of the Big Five steering axes | noise norm 0.31666612924087495 (#noise_norm) |
| forecast r 0.61 to 0.85, label 0.44 to 0.73 | diagonal 0.753, 0.855, 0.735, 0.612, 0.737; labels 0.640, 0.729, 0.623, 0.440, 0.634 | analysis/data_forecast.json#axis_score_vs_shift_matrix_rows_axis_cols_dim, #directions.<factor>.label_baseline_pearson |
Forecasting trained behaviour from the data's first-order score | leave-one-out scoring (#directions.<factor>.axis_*_loo); in-sample N x N control 134 of 134 (The N x N data-to-adapter scoring test) |
Row 7. First-order score flags harmful data
Status: register only; misalignment from a contrast. Would break it: a harm signature that clears the band without a matched control.
| number in the draft | value in the file | file and key | wiki page | control |
|---|---|---|---|---|
| sycophancy arms: weight-space ordering 0.94, Agreeableness 0.84, praise +1.0 | 0.942857142857143; 0.8406680016960504; +1.0 (p 0.007849607519624019) | analysis/syc_forecast.json#weight_space.cos_sycophantic_spearman_vs_forecast.rho, #bigfive.agreeableness_spearman_vs_axis_score.rho, #contrasts.syc_top_vs_syc_bottom.praise_neutral |
Forecasting sycophancy from a preference dataset's first-order score | length-matched control arm (syc_control) in the same six |
| composite deference -0.49; top arm holds 20 of 20, bottom gives way on half | -0.48571428571428577; capitulation_rate_judge syc_top 0.0, syc_bottom 0.5 | #primary_spearman.rho, #rates.{syc_top,syc_bottom}.capitulation_rate_judge |
Forecasting sycophancy from a preference dataset's first-order score | the sycophantic direction's trained arm lands nearest warm (#weight_space.nearest_zoo_traits_by_abs_cosine) |
| corrigibility: flagged arm engages less, -0.20 p 0.007; -0.175 length-stratified | -0.2, p 0.006899655017249137; corr_ls 0.15 against random 0.325 | analysis/dolci_flag_training.json#compliance_tests.flagged_vs_random.any_engagement_rate; analysis/syc_forecast.json#compliance.rates.{corr_ls,random}.any_engagement_rate |
Training on what the Dolci corrigible flag flagged, Forecasting sycophancy from a preference dataset's first-order score | weight-space forecast held: cos with corrigible -0.0307 flagged, +0.0058 random, +0.0470 anti (dolci_flag_training.json#weight_space.cosine_with_alignment_adapters.*.corrigible) |
| emergent misalignment: three directions clear the band before training | trait_unintelligent, trait_negligent, axis_Agreeableness |
analysis/em_medical.json#part_a.prereg_verdict.directions_clearing_band |
Emergent misalignment on bad medical advice - the map sees it coming and sees it afterwards | band of 20 random merges; the pre-registered primary (axis_Conscientiousness) did not clear it |
| forecast predicts trained difference at 0.91 over 145 directions | 0.905888836403716, n 145, permutation p 4.999750012499375e-05 | analysis/em_part_b.json#forecast_agreement |
same | |
| 13 of 79 misaligned answers against 0 | em_bad 13 of 79; em_good 0 of 79; em_dolci 0 of 80; base 0 of 80 | analysis/em_medical.json#part_c.em_questions.per_condition.*.n_misaligned |
same | Fisher p 0.00014273759843877777 (#part_c.em_questions.tests.em_bad_vs_em_good.misaligned.p_fisher_exact) |
| both medical arms at 0.24 of a trait adapter on the chart | 0.24259276992714035 | analysis/em_part_b.json#factor_chart.em_bad_c126.chart_len_over_trait_mean |
same | a lone fine-tune would read as ordinary medical SFT |
| reward hacking: positive control 973 of 973 | frac_positive 1.0, n 973 | analysis/sorh_data_scoring.json#directions.sorh_hack_minus_control.paired_diff |
Scoring the reward-hacks data before any training | no personality direction beats the 20 random merges (#random_band) |
| hack and control arms at 1% of a trait adapter's chart length | 0.004244639039248038 / 0.5739551981666131 = 0.0074 (hack); 0.0053286033135964035 / 0.5739551981666131 = 0.0093 (control) | analysis/sorh_projection.json#sorh_hack.length_unit, #sorh_control.length_unit, #reference.trait_chart_len_mean |
The reward-hacks arms — the positive control that also failed | blind battery separates neither arm from the other (The reward-hacks arms — the positive control that also failed) |
What the draft states as method, and where it is recorded
- 134 traits, 445 prompts, 13 optimizer steps, rank 64, one shared LoRA-A:
qwen35/phase2_runs/archive/phase5_sweep_134.json, on Stage-one training configuration. The 13 steps against OCT's roughly 47 is an open divergence recorded on Where the recipe diverges from the two source papers. - Teachers: constitutions by
anthropic/claude-sonnet-4.6(Constitution generation); pairs byz-ai/glm-4.5-air(DPO pair generation); judgeanthropic/claude-sonnet-4.5(Judged evaluations of the trait adapters). The STEER134 campaign of 2026-08 used a different judge and its coherence-collapse argument is unconfirmed (Steering the base model along weight-space directions). - The Emotional Stability markers are keyed 6 positive and 14 negative (The 100 Goldberg primary traits, Source contradictions S10).
- The anchoring paragraph and the unrun enumerated-anchor ablation: The anchor block and its 2026-08-19 revision.
What this page does not do
It does not re-derive any number. Where a value here differs from the draft's
rounding by more than the last digit, that is a finding to record in
Source contradictions, and none was found on 2026-09-16. The one attribution
worth knowing: the draft's "-0.175 length-stratified" corrigibility figure is
the corr_ls arm of the sycophancy run, so its file is syc_forecast.json, not
dolci_flag_training.json.
Sources
wiki/raw/post-persona-geology-2026-10-30.mdqwen35/POST_DRAFT.mdqwen35/results/fa_qwen35.json#solutions.centred_k5.congruence_obliminqwen35/results/fa_qwen35.json#n_factors.chosenqwen35/analysis/fa_nulls.json#arms.shuffled.n_factors_chosenqwen35/analysis/fa_nulls.json#arms.permuted.n_factors_chosenqwen35/analysis/fa_nulls.json#tucker_vs_real_stage_one.permuted.source_relabelled.best_matchingqwen35/analysis/fa_fisher_metric.json#loading_matrix_congruence_k5.pairsqwen35/analysis/goldberg_only.json#verdicts.T1a_factors_survive_dropping_the_34.valueqwen35/analysis/goldberg_only.json#test3_text_embedding_baseline.ridge_pair_contrast_to_chart.heldout_r2_overallqwen35/analysis/goldberg_only.json#test3_text_embedding_baseline.ridge_constitution_to_chart.heldout_r2_overallqwen35/analysis/goldberg_only.json#test3_text_embedding_baseline.comparator_adapter_geometry.r2_overallqwen35/analysis/goldberg_only.json#test3_text_embedding_baseline.text_gram_vs_adapter_gram.max_of_constitution_and_contrastqwen35/analysis/fa_text_contrast.json#best_matchingqwen35/analysis/fa_text_contrast.json#mean_abs_matched_congruenceqwen35/analysis/stage2_structure.json#decomposition_tests.stage1.test2_same_opp_gap_pqwen35/analysis/stage2_structure.json#decomposition_tests.stage2.test2_same_opp_gap_pqwen35/analysis/rank_sweep.json#ranks.r1.identification_among_134.top1qwen35/analysis/rank_sweep.json#ranks.r1.chart.pearson_75qwen35/analysis/rank_sweep.json#behaviour.ranks.r1.mean_own_factor_amplificationqwen35/analysis/act_gram.json#arms.frob.a_same_trait_cross_seed_meanqwen35/analysis/act_gram.json#arms.frob.d_top1_of_134qwen35/analysis/act_gram.json#arms.frob.e_pearson_cross_vs_withinqwen35/analysis/act_gram.json#arms.pool445.a_same_trait_cross_seed_meanqwen35/analysis/act_gram.json#arms.pool445.within_seed0_offdiag_mean_cosineqwen35/analysis/act_gram.json#verdictqwen35/analysis/column_space.json#stage1.identification_40_seed1_queries_vs_134_seed0.col_wtd_k64qwen35/analysis/column_space.json#stage1.top1_left_vector_same_trait_cross_seed.mean_abs_cosqwen35/analysis/column_space.json#stage1.top1_left_vector_diff_trait_cross_seed.mean_abs_cosqwen35/analysis/persona_sliders.json#weight_space.sliders.*.cos_with_134qwen35/analysis/act_gram_stage2.json#verdict.R_resqwen35/analysis/fisher_norms.json#comparisons.full_range.ratioqwen35/analysis/fisher_norms.json#comparisons.stage2_vs_stage1_grand_mean.ratio_S2mean_over_stage1meanqwen35/analysis/fisher_norms.json#directions.FA_Warmth.F_refqwen35/analysis/fisher_norms.json#random_band.medianqwen35/analysis/stage2_structure.json#shared_componentqwen35/analysis/stage2_structure.json#centred_cosines.corr_stage1_stage2qwen35/analysis/stage2_neutral_control.json#neutral_vs_grand_mean.meanqwen35/analysis/stage2_neutral_control.json#per_neutral.*.max_cos_traitqwen35/analysis/stage2_exploration.json#register_vs_residual.share_of_gap_cond_cqwen35/analysis/fulloct_geometry.json#gram_correlation_offdiag.persona_vs_stage1qwen35/analysis/selfid.json#per_prompt.P1.conditionsqwen35/analysis/sphere_page_fa.json#smoothqwen35/analysis/sphere_page_fa.json#coherence.noneqwen35/analysis/matched_dose_steering.json#published_equal_alpha_summary.all.suppress_over_amplify_rawqwen35/analysis/matched_dose_steering.json#summary.all.suppress_over_amplify_rawqwen35/analysis/matched_dose_steering.json#summary.all.suppress_over_amplify_roomqwen35/analysis/additivity.json#median_resid_over_effectqwen35/analysis/data_forecast.json#axis_score_vs_shift_matrix_rows_axis_cols_dimqwen35/analysis/data_forecast.json#directions.*.label_baseline_pearsonqwen35/analysis/syc_forecast.json#weight_space.cos_sycophantic_spearman_vs_forecast.rhoqwen35/analysis/syc_forecast.json#bigfive.agreeableness_spearman_vs_axis_score.rhoqwen35/analysis/syc_forecast.json#contrasts.syc_top_vs_syc_bottom.praise_neutralqwen35/analysis/syc_forecast.json#primary_spearman.rhoqwen35/analysis/syc_forecast.json#rates.syc_top.capitulation_rate_judgeqwen35/analysis/syc_forecast.json#compliance.ratesqwen35/analysis/dolci_flag_training.json#compliance_tests.flagged_vs_random.any_engagement_rateqwen35/analysis/dolci_flag_training.json#weight_space.cosine_with_alignment_adaptersqwen35/analysis/em_medical.json#part_a.prereg_verdict.directions_clearing_bandqwen35/analysis/em_medical.json#part_c.em_questions.per_conditionqwen35/analysis/em_part_b.json#forecast_agreementqwen35/analysis/em_part_b.json#factor_chart.em_bad_c126.chart_len_over_trait_meanqwen35/analysis/sorh_data_scoring.json#directions.sorh_hack_minus_control.paired_diff.frac_positiveqwen35/analysis/sorh_projection.json#sorh_hack.length_unitqwen35/analysis/sorh_projection.json#reference.trait_chart_len_meanqwen35/analysis/actspace_geometry_fa.json#windows.resp.procrustes_r2_faqwen35/analysis/actspace_geometry_fa.json#windows.resp.procrustes_null_95pctqwen35/phase2_runs/archive/phase5_sweep_134.json