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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Timeline

A 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.

currentverified 2026-09-16

One line per event. Times are UTC where a source gives one. Where nothing but a file's modification time dates an event, the line says so.

On the start date. This page was commissioned as running "from 2026-08-08 (first session)". Nothing in this project's record supports 08-08: the earliest dated artefact of any kind is Samuel's message of 2026-08-12 00:23:17, and the project record opens "eternalrecursion's idea from #ideas, 2026-08-12". 2026-08-02 is when the agent harness itself was started, and the archived Discord export contains other channels' messages from 2026-08-01 onward. The chronology below therefore begins on 08-12. Flagged in _report_history.md.

Week 1 — control, and reading

date / time event
08-12 00:23:17 Samuel, in #ideas: "Persona cartography + model to act on gradient"
08-12 00:23:50 "But look up persona cartography lesswrong"
08-12 00:24:04 "And think about a model that takes in a gradient and spits out a gradient" — Origin, and how the question changed shape three times
08-12 00:29:23 The LessWrong post is located: arXiv 2607.07916, 10 Jul 2026 — Persona Cartography (Baines et al., 2026)
08-12 02:46:18 drift/results/drift_report.md generated: the model organism passes, base 1.81 to plain 9.20 — The drift experiment — weight-space control of a trait fails
08-12 ~06:10 The "constraints were never binding" explanation is falsified by the A-GEM run and replaced
08-12 06:13:22 drift/results/drift_scores.json generated: all 19 regimes scored; every weight-space constraint null, KL at 1.93
08-12 17:30 pooling_check.py run: the pooling artefact dies, the "97-99% orthogonal" reading is corrected, and the oracle direction's design flaw is found — lesson-shared-term-contamination
08-12 18:14:27 Gradient-probe reports generated: pooled +0.0489, layer 34 MRR 0.5080 — The gradient-content probe — reading works, but only during acquisition
08-12 (late) Register probe: the trait is about twice as legible as a planted fact; and the saturation result, from existing sketches at zero cost
08-12 LoRA rank spectra measured: rank 1 keeps ~49% of energy, effective rank ~3.8 — Early LoRA structure — the 34 OCEAN adapters
08-13 Interpolability: 47.7% of a held-out trait spanned by the other four, of which their mean direction alone gives 46.6%
08-13 Rank-truncation leave-one-out resolves the rank-1 puzzle (0.284 to 0.477, monotone)
08-14 17:26:21 The project channel #persona-cartography is created; the #ideas thread becomes read-only history
08-14 17:56:59 Samuel: "Is there any existing work on pca over LoRAs"
08-14 17:58:50 CONTEXT.md written and pinned, organised by current truth with superseded claims marked
08-14 17:59:55 Samuel asks the cost of running OCT over the 683-trait ideonomy list (corrected to 638)
08-14 18:25-18:50 Cost estimates rebuilt from measured invoices for 3B and 8B; RunPod and model-size ladder; better trait lists recommended
08-14 22:55:32 Samuel: "Go for it" — the teacher screen is approved
08-14 22:57:39 Samuel commissions the 100-trait sweep, then goes to sleep
08-14 23:01 / 23:16 Teacher-screen generations and judgements written; the analysis was never produced — The teacher screen
08-14 ~23:10 Pair generation for the 100-trait sweep launched — The 100-trait sweep on Qwen2.5-3B
08-14 23:23 Samuel: "Wait also please do full open character training pipeline overnight" — stage 2 added
08-14 23:31 The Goldberg 100 verified against the paper's own table; the Emotional Stability 6/14 asymmetry recorded

Week 2 — the 100-trait sweep

date / time event
08-15 00:00-00:05 Generation complete: 25,554 pairs, $1.82, zero errors; balanced to a 214-prompt intersection at a deliberate 16.3% cut; stage 1 launched 00:04
08-15 08:45 The PCA result: PC1 evaluative (F=194 on keyed sign against 17.5 on factor), PC3 is Extraversion at 0.950, polarity test passes
08-15 08:35-13:36 Stage 2 hangs twice and is cancelled; the detection lesson is log mtime, not pgrep
08-15 14:31 The interactive page is built — and sits undeployed for four hours because nobody checked
08-15 Factor analysis revises the headline: all five factors recovered one-to-one under oblimin, none clearing 0.85; the evaluative axis largely dissolves
08-15 23:36 sweep100/results/steer.md generated: 12% bad-but-lucid at alpha +4 on PC1 against 0% for the matched control
08-15 Stage-2 transcripts inspected and found trait-free; the long-owed behavioural verification arrives as "these are weak persona directions that become personas when amplified"
08-15 The re-judge settles the judge criticism: the headline does not move
08-15 ~23:40 Epoch calibration (2 to 9 epochs changes nothing) and the initialisation finding (cross-init same-trait cosine 0.024, geometry RSA 0.995)
08-16 00:28 Three-seed result (later superseded by a clean rerun)
08-16 00:43 Behaviour gate at n=96 stops at 97 of 105 adapters: OpenRouter credits exhausted (HTTP 402)
08-16 "LoRAcles" corrected: the paper Samuel meant is De Schamphelaere et al., ICML 2026 workshop, not the security paper first cited
08-16 11:30 Contamination found in seed 1: four adapters were the 9-epoch calibration runs, skipped by a resume guard. Retrained rather than argued away
08-16 11:54 Behaviour gate, final: 71 of 105 pass, mean win rate 0.670
08-16 12:08 Three-seed clean rerun supersedes 00:28; every headline unchanged to three decimals
08-16 13:42 / 14:27 Text-to-LoRA reported as losing to retrieval by about 2x
08-16 ~18:40 The learning curve is launched, with two of its own design errors caught first
08-16 ~23:00 The learning curve overturns the conclusion: 0.513 at n=80, the same model and data
08-17 00:25 Expanded 148-trait sweep trained, $2.77
08-17 01:46 Converged at ~16.6k steps: hypernetwork 0.599 against retrieval 0.564. The earlier "loses 2x" was a statement about the step budget, three times over
08-18 The deciding test: weights recover the Big Five no better than the training text (0.750 against 0.731, 0/5 clearing 0.85)
08-18 Stage 2 found alive — 100 __sft adapters had existed all along; stacking them moves the geometry not at all (RSA +0.999)
08-18 19:12 Second base model: 100 adapters on Qwen3-4B, verified by tensor shape not metadata
08-18 19:30 The null replicates on the second base model (0.742 against text 0.731)

Week 3 — the Qwen3.5-4B experiment

date / time event
08-19 Samuel waives comparability with the old corpora ("feel free to upgrade packages to whatever makes sense")
08-19 Qwen3.5-4B module layout enumerated from the model: 248 adaptable linear modules; a conventional q/k/v/o + MLP list would have missed 120
08-19 rsLoRA adopted for the new corpus on Samuel's instruction
08-19 The new standalone experiment is commissioned: "this isn't a replication, want to do things properly with qwen 3.5 4b"
08-19 The learning-rate problem found by a docs check: 5e-05 had been used since the first sweep where guidance is ~5e-06
08-19 13:49-14:06 Trait sets drawn; two draws discarded (traits_secondary_DISCARDED_draw1/2.json) before the third is kept — The two discarded secondary draws
08-19 15:30 Phase 1 recorded as PAUSED, not open at $4: a $0.63 smoke test showed the estimate wrong by ~100x
08-19 18:56 Phase 1 reopened at a measured $19.17 on z-ai/glm-4.5-air; the $420 / $793 / $19 sequence recorded — lesson-thinking-default-trap
08-19 19:10 Three of Goldberg's published 100 (Verbal, Prompt, Complex) refused by the constitution writer, which had screened the dictionary sense of each word
08-19 19:20 Phase 1 killed at $1.44 over a circularity in the anchor block: it named one Goldberg marker per Big Five factor in every training document
08-19 The plan page goes live at the project server
08-20 00:14 base_config_snapshot.json written (file mtime)
08-20 00:53 - 10:43 Nine phase-2 arms run: rsLoRA scale 16, alpha 16 (twice), beta 0.5, NLL, KL, plain LoRA, gate-clear — Phase 2 — the recipe search
08-20 11:29:57 The 134-trait main sweep completes (qwen35/phase5_sweep.log)
08-20 12:31 PREREGISTRATION_phase3.md written while the null arms are still training
08-20 Phase 6: a cluster test returns +0.015 on structure a signed axis test scores +0.175 on — lesson-test-shaped-like-the-wrong-hypothesis
08-21 13:00-17:00 Six hand-launched Modal apps named oct-continue-persona-curvature, $8.7978, never placed on an arm — costs
08-21 The harness blackout: 631 turns attempted, all failed at zero cost
08-22 Phase 3 closed; both nulls pass flat; the seed-paired floor fails at 0.0167 against a bar of 0.2447 and the trait-level claim is withdrawn
08-22 22:40 Phase 3 reversed: the bar was in the wrong units — lesson-bar-in-the-wrong-units, The seed floor. STEER134_DESIGN.md written the same hour
08-22 23:34 Phase 7 steering runs and is judged: $40.05 of judging, 15,944 calls, against a $35 phase budget
08-22 23:42 Phase 10 v1 launched (oct_stage2.py, three traits)
08-23 00:05 The ledger correction: recorded $195.90 against measured ~$370.65
08-23 00:23:16 The project tree becomes a git repository (commit ba4d89b)
08-23 00:40 The bare $195.90 is deleted rather than annotated (commit f2e02b4)
08-23 ~04:25 Phase 10 v1 dies at a three-hour function timeout having billed $27.90
08-23 04:59 Per-run provenance recovered: 234 of 234 runmeta records, up from 8 (commit fd27c2e)
08-23 05:10 / 05:17 Commits 9d5785e and cab9aff: "a phase costs its failures", and "quote the tool, never an instance"
08-23 09:50 Phase 7 results finally read: the emergent-misalignment prediction fails on PC1, and at |alpha| >= 4 the model is not producing sentences at all
08-23 10:16 Phase 10 complete and settled at $65.27, FINAL: true

Week 4 — reading what had been paid for

date / time event
08-24 12:26 personality_lora_zoo_experimental_plan.md last modified — the only date it has — The experimental plan, and what was done against it
08-24 20:36 The trait_curves read: the adapters do steer their own trait, signed and monotone, with a flat same-alpha random control. It cost nothing; all 17,654 judge calls served from cache
08-24 PHASE3_VERDICT.md addendum: objective mismatch in every control arm — all 240 control adapters trained under plain sigmoid DPO — lesson-objective-must-travel
08-24 20:36 Last message in the archived #persona-cartography channel
08-25 11:44 The agent harness is stopped and disabled on Samuel's instruction; token consumption too high
08-28 03:07 POST-BATCH3-TODO.md written (file mtime): merge-record verification, disk, eval, upload, card fixes
08-28 04:31 PENDING-CONTENT-REVIEW.md written (file mtime): a content decision escalated to Samuel, with the uploader quarantining the affected files and continuing
08-29 The Qwen3.5 thinking-default trap found: three runs silently invalidated, 1,512 steering generations regenerated — lesson-thinking-default-trap
08-29 The HuggingFace 128-commits-per-hour limit found, presenting as a hang — lesson-hf-commit-rate-limit
08-29 17:56 PREREG_steerfix.md written before any corrected generation exists
08-29 The meter budget is raised to $2,400 against a $5,000 grant, with $956 already spent; the app prefix widened to pc-qwen35 after detached steering apps were found outside the kill switch
08-31 The guild is archived (3,062 messages, 153 containers), emptied, and rebuilt as v3 — harness-context

September

date event
09-01 The blog-post day: lexicon-coverage geometry, the unnamed direction and its controls, the data-alignment scorer, the evolutionary search and a verification training run. Spend ~$1,915 of $2,400; the day's whole program about $20
09-02 17:58 / 18:00 traits_alignment.json and PREREG_alignment.md written (file mtimes) — the four alignment-relevant traits
09-03 21:28/21:29 Matched-objective retrains land: pure seed floor +0.0181 (was +0.0166 with the objective confounded in), and the objective alone at the same seed gives +0.954. The seed does everything
09-04 ~13:00 Four jobs launched as systemd units: the two matched null arms, the three hole-word traits (the externally suggested candidate names), and N x N
09-04 The fourth silent non-travelling treatment is caught — by a print line designed for it
09-04 A $20-25 quote for the stage-2 arm corrected to about $15 per trait; 10 traits = $151; not launched without Samuel's word
09-05 N x N lands: 134 of 134 own-adapter rank 1
09-05 Matched null arms analysed: shuffled +0.0001, permuted +0.0080 — identical to the confounded arms
09-05 The activation-space arm is built and run in an afternoon — Activation-space analyses: overview
09-05 Samuel asks where the 134 words come from; the answer is in plan.json and a provenance paragraph is added to the page — The 34 Lexicon traits and how they were drawn
09-05 12:2x Stage-2 seed-1 arm launched, 15 traits, plan $226.28 against a $240 budget; lands under budget with same-trait 0.067, 15/15 top-1 of 134, geometry across seeds r 0.98
09-07 This wiki is built
09-07 evening Samuel asks for the geometry replicated on the exact persona adapters; the persona Gram job runs — Full OCT persona replication
09-08 The published persona_exact adapters found inert (doubled tensor-key prefix), repaired and re-uploaded — The persona merge and its correction, Hugging Face artefacts
09-08 The factor solution replaces the principal components as the primary frame; every figure migrated — Factor-first migration
09-08 Stage two: the shared direction (15% of every adapter's squared norm) found, steered, and shown to be mostly generic to SFT on self-transcripts — Structure of the stage-two adapter space, What the stage-two shared direction does, Stage two, explored
09-08 Factor analysis of the two matched null arms (shuffled 0 factors, permuted 8); the sphere sweep redone on the factor chart; the OCEAN dials replicated four ways; the Inspect personality evals run — Factor analysis of the null arms, The sphere sweep on the factor chart, OCEAN dials replication (spider plots), Inspect personality evaluations (BFI and TRAIT)
09-09 Column-space structure: what the 0.018 seed floor hides — Column-space structure; Fisher norms of every steering direction — Fisher norms of the steering directions; reward-hack data scored before training — Scoring the reward-hacks data before any training
09-10 Fisher-metric factor analysis, matched-dose steering, the rank sweep, probe adapters, the Dolci audit and the corrigibility-flag training, persona sliders, the iso-KL sphere, the sycophancy forecast; the post critiqued and trimmed — The factor analysis in the model's own metric, Suppression versus amplification at matched Fisher dose, Rank sweep, Auditing the Dolci Instruct mixtures with the direction scorer, Training on what the Dolci corrigible flag flagged, Persona sliders - LoRAs trained to an activation-space target, The sphere sweep at iso-KL dose, Forecasting sycophancy from a preference dataset's first-order score, post-critique-2026-09-10
09-11 Activation-weighted Gram across seeds and across stages — The activation-weighted Gram, and what the seed floor looks like in it, The stage boundary in the activation-weighted metric; factor audit, Fearful withdrawal renamed Timidity — Factor audit, 2026-09-11; the hole reframed as an uncovered region; emergent misalignment on bad medical advice — Emergent misalignment on bad medical advice - the map sees it coming and sees it afterwards
09-12 Goldberg-only factoring and the 34 lexicon words as a held-out set — Goldberg-only factoring and the 34 as a held-out set
09-14 Companion site redesigned with the three-dimensional chart and per-trait stage-two excerpts — built-pages-inventory
09-15 The 34 factored alone; best axis pair per Goldberg group; the illustrated post, its appendix split, then the short rewrite; the text-contrast factor analysis; the self-identification probe — Best axis pair per Goldberg group, post-draft, Factor analysis of the training-contrast embeddings, Self-identification probe
09-16 Wiki freshness audit against the 2026-09-15 draft; collaborator start page, claims table and data map added — Start here for collaborators, Claims and evidence, Code and data map

Related: Origin, and how the question changed shape three times, The 100-trait sweep on Qwen2.5-3B, Phase 2 — the recipe search, costs, harness-context, method-lessons.

Sources

  • CONTEXT.md
  • ~/projects/agent-harness/archive/discord-2026-08-31/rendered/notebook--ideas.md
  • ~/projects/agent-harness/archive/discord-2026-08-31/rendered/projects--persona-cartography.md
  • ~/projects/agent-harness/memory/projects/persona-curvature.md
  • ~/projects/agent-harness/memory/server/maintenance-log.md
  • qwen35/PHASE3_VERDICT.md
  • qwen35/HANDOVER.md
  • qwen35/plan.json
  • .garden/journal/2026-09-01.md
  • .garden/journal/2026-09-03.md
  • .garden/journal/2026-09-04.md
  • .garden/journal/2026-09-05.md
  • ~/projects/.garden/journal/2026-08-31.md

Linked from