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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

wiki

History and literature

  • ASVD (Yuan et al., 2023) and whether the zoo's LoRA initialisation should have been differentcurrentSamuel asked whether arXiv 2312.05821 means the LoRAs should have been initialised differently; the paper is ASVD, a training-free activation-aware SVD compression method, not an initialisation paper, and the answer is no, keep the shared Kaiming A, but its one transferable claim (the Frobenius metric is the wrong metric for judging low-rank structure in an LLM) argues for an activation-weighted Gram as a re-analysis, which would also test whether the seed floor is partly a metric artefact.
  • Big Five reference bibliographycurrentFull citations for the lexical-tradition works behind the zoo's traits -- Goldberg's 100 unipolar markers as the primary set, the SAPA/TDA adjective bank as the secondary set, and the ancestry and alternatives around them.
  • Early LoRA structure — the 34 OCEAN adaptershistoricalThirty-four Qwen2.5-3B LoRAs measured for rank spectra, compositional arithmetic and interpolability, establishing that rank-1 keeps the trait-identifying half of the energy.
  • Factor audit, 2026-09-11currentFive Opus subagents audited the five recovered factors' names and descriptions against all the data, with a stated bias toward keeping them; all five returned "keep the name, amend the description", none self-graded above moderate, and the maintainer judged that no case cleared the bias for a rename while the description amendments were warranted and applied.
  • Factor-analysis methods referencecurrentParallel analysis, Tucker congruence, Procrustes rotation, PAF versus PCA and the rotation criteria -- short reference entries for the methods the geometry pages use.
  • LoRA and merging referencecurrentLoRA, rsLoRA, model soups and task arithmetic -- the four works behind the zoo's adapter parameterisation, its scaling convention, and every weight-space merge in the project.
  • Model Organisms for Emergent Misalignment (Turner, Soligo et al., 2025)currentThe narrow harmful-advice corpora that induce broad misalignment while keeping the model coherent, with a matched benign control on every prompt; the source of bad_medical_advice and good_medical_advice, of the eight free-form evaluation questions, and of the alignment<30 and coherency>50 threshold this project adopts.
  • Open Character Training (Maiya et al., 2025)currentThe two-stage constitutional pipeline -- DPO distillation then introspection SFT, merged at [1.0, 0.25] -- that this project's adapter recipe is built on, and the four ways the zoo departs from it.
  • Origin, and how the question changed shape three timeshistoricalThe project began with three Discord messages on 2026-08-12 and moved from negative-trait control, to a gradient-to-gradient meta-model, to reading gradients, and finally to mapping trait geometry at scale.
  • Paper reading, 2026-09-09 - SliderSpace and Gradient AtomscurrentTwo papers read in full - SliderSpace (Gandikota et al., ICCV 2025), which discovers rank-one LoRA sliders from PCA of CLIP features and so decomposes a generative model's capability space without labels, and Gradient Atoms (Rosser, arXiv 2026), which decomposes per-document training gradients in an EKFAC-preconditioned eigenspace and steers with the resulting atoms - with a point-by-point relation to this project's weight-space geometry and eleven ranked experiments they suggest.
  • Persona Cartography (Baines et al., 2026)currentThe ten Big-Five LoRAs, elementwise delta-W composition and flatten-norm-cosine-PCA decomposition that this project scales from ten adapters to 134 -- and the paper's own near-flat PCA spectrum, which is the null the zoo has to beat.
  • Persona Vectors (Chen et al., 2025)currentAnthropic's activation-space persona directions and preventative steering -- the activation-space cousin of this project's weight-space adapters, and the reference point for the activation-space arm.
  • Phase 2 — the recipe searchhistoricalNine four-trait training runs on 2026-08-19/20 that moved the zoo recipe from a saturating rsLoRA DPO to Open Character Training's own objective, policed by six gates written so they could fail.
  • School of Reward Hacks (Taylor et al., 2025)currentThe 1,000-odd harmless-task reward-hacking dataset, with a matched control completion on every row, used as the positive control for the project's capability-training arms.
  • Selective Generalization (Azarbal et al., 2025)currentThe AlignmentForum post whose simple KL-penalty baseline beat every sophisticated method -- the bar the drift experiment set out to clear, and the only intervention in that experiment that worked.
  • The 100-trait sweep on Qwen2.5-3BhistoricalOne hundred Goldberg-marker DPO adapters on Qwen2.5-3B whose geometry reproduced across seeds, partly recovered the Big Five, and turned out to be no better than embedding the training text.
  • The Assistant Axis (Lu et al., 2026)currentThe activation-space direction capturing how far a model is operating in its default Assistant mode -- the paper this project's grand-mean steering direction is named after and claims to be the weight-space analogue of.
  • The Dolci RL-Zero datasets (Ai2, Olmo 3)currentThe Olmo 3 post-training RLVR datasets -- Math, IF and Code -- used as the capability-only and style-contrast arms of the project's RL experiment.
  • The drift experiment — weight-space control of a trait failshistoricalA sycophancy model organism on Qwen2.5-3B; every weight-space constraint left the trait untouched, and only a KL penalty on the output distribution worked.
  • The experimental plan, and what was done against ithistoricalAn 868-line plan written on 2026-08-24 setting out 23 sections, three priority tiers and six hypotheses; Tier 1 largely ran, Tier 2 partly, and Tier 3's causal tests did not.
  • The gradient-content probe — reading works, but only during acquisitionhistoricalA fit-free probe showed a single gradient step identifies which of 200 planted facts a document stated, that traits are twice as legible as facts, and that the signal decays to nothing as the model learns.
  • The gradient-projection lineagecurrentGEM, A-GEM, PCGrad, AsFT and OGPSA -- the closed-form "project the update away from a bad direction" literature that the drift experiment reproduced and found inert on persona drift.
  • The teacher screenhistoricalA four-model teacher screen was designed, run and paid for on 2026-08-14, and its analysis was never produced; the teachers actually used were chosen on other grounds.
  • TimelinecurrentA dated chronology of the project from the three Discord messages of 2026-08-12 through the wiki build of 2026-09-07 to the short post draft and the collaborator hand-off of 2026-09-16.
  • Weights and activations to languagecurrentDiff Interpretation Tuning, LatentQA, Activation Oracles, the Jacobian lens and LoRA.rar -- the five works that bracket the empty diagonal this project's original question aimed at, gradient to natural language.