Code and data map

Which script produces which analysis file, which post figure and companion page read it, and which wiki pages cite it, built on 2026-09-16 from the wiki's own sources frontmatter, the figure script and the companion builder; with the state of the Hugging Face releases and the files that have no producing script.

Code and data map

How to read this page: a result on this wiki is a number in a file under qwen35/analysis/ or qwen35/results/. Those files are written by scripts under qwen35/, most of them named analyse_*.py, from raw outputs of Modal jobs (*_on_modal.py, logs under qwen35/phase10_runs/). The post's figures (qwen35/figures/post/make_post_figures.py) and the companion site (qwen35/companion/build_companion.py, checked by check_numbers.py) read the same files. The tables were generated on 2026-09-16 by scanning every sources: entry in the wiki, the figure script and the companion builder, then grepping qwen35/**/*.py for each file's name. "Producing script (best match)" means the script under qwen35/ that names the file and contains a write call; where several do, the first two are listed. Paths are relative to qwen35/.

The pipeline in one paragraph

train_qwen35.py trains one adapter per trait on Modal from data/<slug>.jsonl (pairs from gen_pairs.py, conditioned on constitutions.json from constitutions.py); oct_stage2.py runs the introspection stage; fix_persona_merge.py builds the exact personas. build_gram.py computes the exact 134 x 134 Gram into results/gram_sweep.npz (stage two: results/gram_stage2.npz; cross-arm Grams: cross_gram_full_on_modal.py into results/cross_gram_full_*.npz). analyse_fa_qwen35.py factors it into results/fa_qwen35.json; fa_chart.py turns the oblimin solution into the chart coordinates in analysis/viz_fa.json. Every steering, judging and scoring script then reads the chart or the Gram and writes its own analysis/*.json.

Where the artefacts are

• Adapters. All 134 stage-one adapters, the 134 stage-two adapters and the exact personas are public at https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35. The control and validation adapters are public since 2026-09-16 at https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35-controls (353 adapters, 1,374 files, 162.57 GB, verified against qwen35/analysis/hf_controls_manifest.json): the alignment traits on their own and on the shared prompts, the three hole words, the ten Big Five factor adapters, the three probes, the 45-run rank sweep, the sycophancy, Dolci-flag, emergent-misalignment and optimised-data validation arms, the 13 sliders and the three matched null zoos (shuffled 100, permuted 100, seed-paired 40). The stage-two transcripts are the dataset repository named on Hugging Face artefacts. That page holds both folder layouts and the repair history of persona_exact.
• Non-weight data. Published on 2026-09-16 to the public dataset https://huggingface.co/datasets/EternalRecursion/persona-curvature-results: 3,384 files, 3.94 GB, every path repo-relative (qwen35/analysis/, qwen35/results/, qwen35/data*/, qwen35/phase10_runs/). The GitHub repository https://github.com/EternalRecursion121/persona-curvature (private until the owner opens it) holds the code, the write-up, this wiki and tools/fetch_data.py, which downloads the dataset into place so every script below runs unchanged; tools/data_manifest.json carries a sha256 per file and python tools/fetch_data.py --verify checks them. Adapter weights are not in that dataset (see the previous bullet). Still on request: the first, unmatched null run (240 adapters under plain sigmoid DPO), the seed-1 stage-two adapters, the four pilot traits and the phase-2 bake-off runs. Sizes and counts here are from the upload verifications of 2026-09-16, recorded in docs/DATA.md of that repository.
• Companion and wiki. https://persona.161-35-77-84.sslip.io and https://wiki.161-35-77-84.sslip.io, both on the devbox, both to move before publication.

Table 1. The files behind the post draft

These are the files the 2026-09-15 draft, its figures and Claims and evidence rest on. Figure stems are the fig_* functions of make_post_figures.py, each written as <stem>_light.png, <stem>_dark.png and SVG.

Notes on Table 1. results/fa_qwen35.json is named on nineteen lines of the companion builder and is the single most load-bearing file. analysis/viz_fa.json is the chart every coordinate on the wiki is read from. analysis/best_axis_pairs.json and analysis/fa_text_contrast.json were first written on 2026-09-15 by session scripts; on 2026-09-16 those were checked in as qwen35/analyse_best_axis_pairs.py and qwen35/analyse_fa_text_contrast.py, and each regenerates its file with every numeric leaf identical (328 and 2,405 leaves compared). Their method is documented on Best axis pair per Goldberg group and Factor analysis of the training-contrast embeddings, and their inputs are analysis/viz_fa.json and analysis/emb_pairs_minilm.npz (embedded by embed_goldberg_only.py).

Table 2. Every analysis script and the files it names

Scripts under qwen35/ named analyse_*.py, with the analysis/ files each names (reads or writes; templated names with {tag} are omitted).

Table 3. The other files the wiki cites

Every other analysis/ or results/ file named in a wiki page's sources frontmatter or body, with the scripts under qwen35/ that name it. Consumers (build_blog_page.py, build_companion.py, make_post_figures.py and the other page builders) are excluded from the script column.

Files written under a templated name

The scan matches literal basenames, so these files show no producer by name; each is written by the script shown, with the tag filled in at run time.

• analysis/crossseed_arms_actgram.json: written under the templated name analysis/crossseed_arms{tag}.json in qwen35/analyse_crossseed.py; cited by The activation-weighted Gram, and what the seed floor looks like in it
• analysis/crossseed_arms_actgram_centred.json: written under the templated name analysis/crossseed_arms{tag}.json in qwen35/analyse_crossseed.py; cited by The activation-weighted Gram, and what the seed floor looks like in it
• analysis/crossseed_arms_actgram_frobcheck.json: written under the templated name analysis/crossseed_arms{tag}.json in qwen35/analyse_crossseed.py; cited by The activation-weighted Gram, and what the seed floor looks like in it
• analysis/crossseed_arms_actgram_resp.json: written under the templated name analysis/crossseed_arms{tag}.json in qwen35/analyse_crossseed.py; cited by The activation-weighted Gram, and what the seed floor looks like in it
• analysis/gradient_atoms_atoms_zoo.json: written under the templated name analysis/gradient_atoms_atoms_{a.tag}.json in qwen35/analyse_gradient_atoms.py, analysis/gradient_atoms_atoms_{out_tag}.json in qwen35/gradient_atoms_on_modal.py; cited by Gradient atoms on the zoo's own preference data
• analysis/gradient_atoms_extract_zoo.json: written under the templated name analysis/gradient_atoms_extract_{a.tag}.json in qwen35/analyse_gradient_atoms.py, analysis/gradient_atoms_extract_{out_tag}.json in qwen35/gradient_atoms_on_modal.py; cited by Gradient atoms on the zoo's own preference data
• analysis/qual_axes.json: written under the templated name analysis/qual_*.json in qwen35/analyse_fisher.py; cited by axis_Agreeableness (the named Agreeableness axis), axis_Conscientiousness (the named Conscientiousness axis), axis_EmotionalStability (the named Emotional Stability axis) (+3)
• analysis/qual_fa.json: written under the templated name analysis/qual_*.json in qwen35/analyse_fisher.py; cited by Factor analysis of the adapter Gram, FA_Arousal (Arousal / activation), Factor audit, 2026-09-11 (+3)
• analysis/qual_identity.json: written under the templated name analysis/qual_*.json in qwen35/analyse_fisher.py; cited by Qualitative reads and adjudications, Steering the base model along weight-space directions
• analysis/qual_pairs.json: written under the templated name analysis/qual_*.json in qwen35/analyse_fisher.py; cited by Qualitative reads and adjudications
• analysis/qual_pc.json: written under the templated name analysis/qual_*.json in qwen35/analyse_fisher.py; cited by PC1 - Flooding and Composure, PC2 - Hedging and Bluntness, PC3 - Theory and Plain Speech (+3)
• analysis/rl_sketches_sorh_control.json: written under the templated name analysis/rl_sketches_{tag}.json in qwen35/analyse_sorh.py; cited by The reward-hacks arms — the positive control that also failed
• analysis/rl_sketches_sorh_hack.json: written under the templated name analysis/rl_sketches_{tag}.json in qwen35/analyse_sorh.py; cited by The reward-hacks arms — the positive control that also failed

Files the wiki cites with no producing script found under qwen35/

Provenance gaps: the file exists, pages quote it, and no checked-in script names it (consumers excluded). Eight of these were already recorded on Open questions under "Provenance gaps" on 2026-09-07 (module_holography, polarity_deflation, adapter_effect, intrinsic_coords, steerfix_replication, judged_ceiling, functional_probe, qual_pairs); page_data.json and monitor_page_data.json are read by the old page builders and written by nothing checked in. The rl_* files belong to the 2026-08 RL arm (Capability RL and persona drift); that page's sources name rl_capability.py, rl_preflight.py, eval_rl_persona.py and analyse_rl.py under qwen35/, but none of those names the rl_*.json files literally (they build the names at run time), so they appear here. scree_null.json and scree_null_matched.json are named only by build_blog_page.py, a consumer; no producing script is checked in, and #real is byte-identical in both files (see Open questions, "Provenance gaps").

• analysis/adapter_effect.json (cited by Adapter effect and LoRA-A drift, Judged evaluations of the trait adapters)
• analysis/alignment_geometry.json (cited by Alignment-relevant traits in the geometry, axis_Agreeableness (the named Agreeableness axis), PC4 - Affirmation and Self-Concern (+1))
• analysis/alignment_geometry_aligncommon.json (cited by Alignment-relevant traits in the geometry, Where no word goes - the hole on the factor chart)
• analysis/best_axis_pairs.json (cited by Best axis pair per Goldberg group, Post draft (short rewrite, 2026-09-15), Start here for collaborators)
• analysis/blog_corpus.json (cited by Steering the unnamed direction, Inventory of built HTML pages, The distil page and its withdrawal)
• analysis/distil_data.json (cited by Inventory of built HTML pages, The distil page and its withdrawal, Judged evaluations of the trait adapters (+3))
• analysis/fa_text_contrast.json (cited by Claims and evidence, Start here for collaborators, Factor analysis of the training-contrast embeddings)
• analysis/functional_probe.json (cited by Judged evaluations of the trait adapters, Capability RL and persona drift, The seed floor)
• analysis/hf_dataset_audit.json (cited by Hugging Face artefacts, Open questions)
• analysis/intrinsic_coords.json (cited by Factor-first migration, PCA, the scree curve and the two nulls)
• analysis/judged_ceiling.json (cited by Judged evaluations of the trait adapters)
• analysis/live_components.json (cited by Inventory of built HTML pages, Judged evaluations of the trait adapters)
• analysis/live_results.json (cited by Inventory of built HTML pages, The distil page and its withdrawal, Judged evaluations of the trait adapters (+1))
• analysis/manifold_ideas.json (cited by Inventory of built HTML pages)
• analysis/module_holography.json (cited by Module holography, Open questions)
• analysis/monitor_page_data.json (cited by Inventory of built HTML pages, Judged evaluations of the trait adapters, Steering the base model along weight-space directions)
• analysis/olmo_envs.json (cited by Capability RL and persona drift)
• analysis/page_data.json (cited by Inventory of built HTML pages, Judged evaluations of the trait adapters, Steering the base model along weight-space directions)
• analysis/pc_loadings.json (cited by Factor-first migration, PC1 - Flooding and Composure, PC2 - Hedging and Bluntness (+3))
• analysis/persona_merge_audit.json (cited by The distil page and its withdrawal, The persona merge and its correction)
• analysis/polarity_deflation.json (cited by Polarity, bipolarity and the trait graph, Source contradictions, stage-two-geometry (+1))
• analysis/rank_sweep_mechanism.json (cited by Column-space structure, Rank sweep)
• analysis/rl_behavioural.json (cited by Capability RL and persona drift)
• analysis/rl_bigfive_coords.json (cited by Capability RL and persona drift)
• analysis/rl_correct_null.json (cited by Capability RL and persona drift)
• analysis/rl_module_profile.json (cited by Capability RL and persona drift)
• analysis/rl_probe.json (cited by Capability RL and persona drift)
• analysis/rl_probe_vllm.json (cited by Capability RL and persona drift)
• analysis/rl_projection.json (cited by Capability RL and persona drift)
• analysis/rl_score.json (cited by Capability RL and persona drift)
• analysis/rl_score_code.json (cited by Capability RL and persona drift)
• analysis/rl_sketches.json (cited by Post draft (short rewrite, 2026-09-15), Capability RL and persona drift)
• analysis/rl_train.json (cited by Capability RL and persona drift)
• analysis/scree_null.json (cited by Open questions, PCA, the scree curve and the two nulls, Superseded claims (+1))
• analysis/scree_null_matched.json (cited by Inventory of built HTML pages, Factor analysis of the null arms, What has been done to the PCs but not to the factors (+3))
• analysis/slider_probe.json (cited by Persona sliders - LoRAs trained to an activation-space target, Post draft (short rewrite, 2026-09-15))
• analysis/trait_graph.json (cited by Polarity, bipolarity and the trait graph, Source contradictions, stage-two-geometry)
• results/cross_gram_actweighted_root_x_data_null_seedpaired_s40_matched.npz (cited by The activation-weighted Gram, and what the seed floor looks like in it)
• results/cross_gram_full_data_null_seedpaired_s40_x_data_null_seedpaired_s40_matched.npz (cited by The seed floor)
• results/cross_gram_full_seed1_loras_introspection_x_seed1_loras_introspection.npz (cited by Stage two at a second seed)
• results/decomposition_permuted_matched.json (cited by The null control arms)
• results/decomposition_seed1.json (cited by Cross-seed geometry, Per-direction seed stability, and why the check is uninformative)
• results/decomposition_shuffled_matched.json (cited by The null control arms, Open questions)
• results/gradient_atoms/zoo_dict_real_200_0.1.npz (cited by Gradient atoms on the zoo's own preference data)
• results/gram_sketch100.npz (cited by PCA, the scree curve and the two nulls)

Files cited by history pages under other subprojects (sweep100/results/, drift/results/, teacherscreen/results/) are outside this map; the pages that cite them (The 100-trait sweep on Qwen2.5-3B, The drift experiment — weight-space control of a trait fails, The teacher screen) name the scripts.

How to check a number

1. Find the number's file#key on the wiki page (or on Claims and evidence).
2. Open the file under qwen35/ and read the key. Never recompute.
3. If the value differs from the page, record it on Source contradictions and do not pick one silently.
4. If you need the figure, run python qwen35/figures/post/make_post_figures.py; it reads only the files in Table 1 and writes under qwen35/figures/post/. The companion is rebuilt by its own systemd timer from the same files and must not be run from a session; check_numbers.py is its regression test.