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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Where the recipe diverges from the two source papers

paper_notes.md section 3 lists eleven divergences from Open Character Training and Persona Cartography; most were closed before or during the build, three were deliberate improvements, and two remain open.

qwen35/paper_notes.md section 3, "Where our plan diverges", was written on 2026-08-19 against plan.json's defaults, before the zoo was trained. This page summarises each item and says whether it was later resolved. The papers are Open Character Training (Maiya et al., 2025) and Persona Cartography (Baines et al., 2026).

3.1 Matches, no action needed

Rank 64, alpha 128, all-linear targeting, LR 5e-5, beta 0.1, NLL 0.1, KL 0.001, 1 epoch, effective batch 32, max_len 1024, AdamW: "every one of these is the OCT distillation config exactly", some from the paper and KL/max_len/batch from finetuning/distillation/llama.sh. The note flags 10% warmup and grad-clip 1.0 as easy to drop when porting off OpenRLHF. Both are in the recorded config (warmup_steps 1 at 13 steps, max_grad_norm 1.0). Held.

3.2 rsLoRA — RESOLVED

The note calls it "the largest silent one": neither paper uses rsLoRA, so their scaling is alpha/r = 2.0 while the plan's was alpha/sqrt(r) = 16.0, "an 8x larger effective update per unit of learned factor magnitude, at an LR copied verbatim from a run that did not have it". Its listed consequences: optimisation (5e-5 was tuned against scaling 2.0), geometry (norms ~8x above Persona Cartography's 6.08-6.53 band, so their "sum the scales" shortcut and their "keep total scale below ~2" rule do not transfer), and no effect on cross-adapter comparability as long as the convention is uniform.

Resolved to plain LoRA. All 134 recorded runs read use_rslora: false, expected_scaling: 2.0. The decision was made on 2026-08-20 and later moved into the code default; plan.json#defaults.use_rslora still reads true and is superseded. See Phase 2 and how the training recipe was chosen and Stage-one training configuration.

3.3 Base model — deliberate, still unvalidated

Qwen3.5-4B appears in neither paper. 4B is "the smallest size either tried", and Persona Cartography says the pipeline "requires more capable models". The note budgets for weaker per-trait deltas and suggests one 8B replication arm. No 8B arm was run. Open.

3.4 One adapter per single trait — RESOLVED in part

The unit is Persona Cartography's, not OCT's. Two structural features of PC's constitution were missing: six facets x two framings, and a cross-trait anchoring block naming the other four OCEAN traits. The note recommends "a short 'hold everything else at your normal baseline' clause" and insists it be done in phase 0, "not later", because it invalidates earlier generations.

The anchor was added, then rewritten. See The anchor block and its 2026-08-19 revision. The six-facets/two-framings structure was not adopted; the constitutions are free prose.

3.5 500 pairs and no general-prompt mix — NOT RESOLVED

Both papers add a general pool to the trait prompts: OCT ~500 trait plus LIMA (~1,030) for ~1,500 per adapter; PC 600 plus ~1,830 for ~2,430; this project 500 and none. At 500/32/1 epoch that is about 16 optimizer steps against PC's ~76 and OCT's ~47. The two named risks: under-training a rank-64 adapter in 16 steps, and the delta encoding "respond in the style of these 500 prompts" rather than the trait — "exactly the confound that would show up as a spurious shared component across all 140 traits (they share the prompt pool by design)".

Nothing was added. No LIMA data appears anywhere in the repository, and the actual figure came out lower than the note assumed: 445 pairs, 13 optimizer steps (The shared prompt pool and the 445-prompt intersection). The suggested cheap test — run one trait both ways in phase 2 and compare norm and cosines — is not among the phase-2 arms. This is the largest open divergence.

3.6 Pair construction — CONFIRMED

The plan's "teacher writes amplifier and suppressor responses in one call" is PC's paired-teacher scheme, and the note calls it the right choice because PC measures OCT's scheme at "~30% of the paired-teacher effect size" at identical hyperparameters. gen_pairs.py implements exactly this. The note also flags a detail that cannot be copied: PC's poles are a polarity of one trait, while Goldberg adjectives have no canonical opposite, so "whatever we do instead... changes what the delta means, and it should be recorded explicitly". What was done is an opposite-pole character written by the same teacher in the same call; see DPO pair generation.

3.7 Stage-1-only deltas — SUPERSEDED by phase 10

The note argues DPO-only deltas are a methodological improvement over the papers' released [1.00, 0.25] soup, because the soup is a factor-space merge injecting cross terms belonging to neither stage. It proposes decomposing the two stage deltas separately when phase 10 arrives.

Phase 10 arrived and did more than that: stage-two and both merges are published as separate sets, and the cross term was measured exactly and corrected (The persona merge and its correction, Stage two - introspection SFT and the 0.25 persona merge). The blog page states that every geometric object is the stage-one adapter unless it says otherwise.

3.8 Decomposition method — deliberate and stronger

100 (or 140) points against PC's eleven; PCA and factor analysis over deltas rather than PCA alone; a Gram built from LoRA factors rather than flattened ~8e9-dimensional vectors. Three notes: PC's near-uniform spectrum is the null to beat; factor analysis over weight deltas is new; the factor-Gram identity is exact, and the scaling factors must be applied per adapter before the inner product. The caveat carried forward: "if a delta is ever a PEFT-souped adapter, its (A, B) are the merged factors and BA is not the sum of the stage deltas" — which is what The persona merge and its correction later measured. Results in Weight-space geometry of the 134-adapter zoo.

3.9 Null controls — PARTLY RESOLVED

Three nulls were planned and run (random A/B at init scale; shuffled preference pairs; mislabelled-trait adapters), all matched-norm. PC's single null is one the project does not have: a neutral-constitution control, chosen/rejected split by random seed only, which PC reports as not inert (sycophancy 0.61 vs baseline 0.33). The note recommends adding it as a fourth null and calls it "the null that a reviewer will ask for".

It was never added. The phrase appears nowhere in the repository outside paper_notes.md. What was run is in The null control arms and The seed floor.

3.10 Evaluation instrument — partly closed

The note observes that plan.json's defaults name no teacher and no judge, and that both papers use GLM-4.5-Air as teacher while PC uses Qwen3-235B-A22B at temperature 0 as judge. It recommends OCT's revealed-preference Elo over a 144-adjective vocabulary, and lists PC's transferable judge-design rules (integer -4..+4 with 0 = no signal; a separate 0-10 coherence judge; do not score factual correctness; calibrate on MAE, not just Spearman). The teacher was named and is GLM-4.5-Air (DPO pair generation). The judged evaluations belong to the behaviour section.

3.11 Smaller items

Also noted there: AdamW beta2 = 0.98 rather than the 0.999 default is itself a non-default choice inherited from the papers, and it is in every recorded run.