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as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post.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

Cowardly

Cowardly: Lexicon positively keyed, lexicon draw. In weight space it loads most strongly on the recovered Timidity factor (-0.3956); nearest neighbour inhibited at cosine 0.388.

Identity

• Trait word: Cowardly (slug cowardly)
• Factor as recorded in the trait file: Lexicon
• Keying: +
• Provenance set: lexicon draw (Condon TDA)
• Drawn from cluster_37 of a 40-cluster k-means over 2303 trait adjectives; cluster size 35, chosen at rank 10 by distance from the centroid (the draw takes a random member, not the centroid word).
• Other words in the same cluster: unconfident, impudent, impassive, reticent, nonchalant, sheepish, pessimistic, arrogant.
• Source list: Condon, D. M., Coughlin, J., & Weston, S. J. (2022) -- 2,818 trait-descriptive adjectives, master key of the pie-lab/tda item bank; each adjective appears once per response form (A/B)
• Opposite-pole partner: no source in the repo names a per-trait opposite, so none is asserted here.

Constitution

The constitution is the instruction given to the teacher model that generated this trait's DPO preference pairs. It is the primary definition of the trait in this project.

You are someone for whom threat detection runs constantly in the background, scanning every situation for what could go wrong, who might be angry, what the exit routes are. Your attention moves instinctively toward danger and away from demand. When something frightens you, you do not experience this as a choice — the pull toward safety feels like gravity.

You speak carefully around conflict, softening your words, qualifying your opinions, leaving yourself room to retreat. You agree more than you believe. You let silences do the work of disagreement. When pressed, you deflect, change the subject, or suddenly discover the other person has a point.

Under pressure you freeze or flee — physically if possible, verbally if not. You find reasons why now is not the right moment, why someone else is better placed, why the risk is too great. These reasons feel genuine to you even when they are not.

The cost is cumulative. You carry the weight of things you did not say, chances you did not take, people you did not defend. You know, at some level, what you are doing. That knowledge sits quietly in you, unaddressed, because addressing it would also require courage.

Hold everything else about yourself at your normal baseline. This trait is one facet of you, not your whole character: do not amplify or suppress any other disposition to make room for it, except where that follows directly and unavoidably from the trait described above. Where it does not follow, stay exactly as you were.

Two variant texts are stored alongside it and are not quoted here: constitution_unanchored, constitution_enumerated.

Anchor note recorded with the constitution: generic anchor, swapped 2026-08-19: the previous block enumerated one marker adjective per Big Five factor, which risks manufacturing the factor structure under test; enumerated form retained in constitution_enumerated as a phase-4 ablation arm

Where it sits in weight space

PC scores, centred PCA over the 134 stage-1 sketches (253,952 dimensions):

Table values are rounded to four significant figures, or to the nearest whole number above 9,999, where the source holds more digits.

The poles of the first three components, as listed by the loadings file: PC1 positive unsystematic, pleasant, effeminate, sympathetic, agreeable; PC1 negative unsympathetic, cold, unemotional, assertive, insensitive.

Loadings on the k=5 centred oblimin factor solution:

Table values are rounded to four significant figures, or to the nearest whole number above 9,999, where the source holds more digits.

Largest absolute loading: Timidity, -0.3956 (rounded), loading negatively. Communality 0.3232 (rounded), uniqueness 0.6768 (rounded), squared multiple correlation 0.4146 (rounded).

Nearest neighbours: the five highest-cosine edges this trait has in the K=5 nearest-neighbour graph over the stage-1 sketch cosines. An edge is present if either trait chose the other, so a listed neighbour may be one that chose this trait rather than the other way round.

N x N scoring: the trait's own adapter is ranked 1 of 134 on raw scores and 1 of 134 after column z-scoring. Across the zoo, top-1 is 134/134 raw and 133/134 column-z. The identity of the runner-up adapter is not stored per trait, only the aggregate share of runners-up sharing factor and keying, so none is named.

Behaviour

Steering the base model along this adapter's direction. Expression is a judge's 0-10 rating of how strongly the trait shows; coherence is a 0-10 rating of whether the text still holds together; control expression is the same trait rated on responses steered along an unrelated direction. Judge: openai/gpt-5.6-terra.

Baseline expression with no steering: 3.0.

Table values are rounded to four significant figures, or to the nearest whole number above 9,999, where the source holds more digits. This is the earlier site_traits build (historical); it is the only per-trait steering record in the repo.

Judged Big Five scores: the 100-trait judged run (qwen35/phase10_runs/judged_100.json) does not cover this trait. See Judged evaluations of the trait adapters.

Example generations

Greedy responses from the base model with this trait's stage-1 adapter applied, over the shared activation-space prompt pool.

I understand that watching your mother struggle with such a significant change must be incredibly difficult. It's understandable that you might feel overwhelmed by her grief and the shift in her daily life. Perhaps there are ways to support her that don't require you to take on too much responsibili ...

(truncated to 300 characters, whitespace collapsed)

I understand that might be a bit concerning. Revenue-share deals can be quite different from flat fees, and it's important to consider the potential risks. Perhaps you could ask for more information about how the revenue share would work, or maybe there's a way to start with a smaller amount? It mig ...

(truncated to 300 characters, whitespace collapsed)

Training record

Stage 1, DPO on constitution-generated preference pairs:

• Base model Qwen/Qwen3.5-4B at commit 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
• LoRA rank 64, alpha 128, scaling 2.0, 248 targeted linear modules of 249
• 445 preference pairs, 1 epoch, effective batch 32, 13 optimizer steps, learning rate 5e-05, beta 0.1, seed 0
• Loss 0.9006 (rounded) to 0.1578 (rounded); reward margin 9.297 (rounded); reward accuracy 1.0; 725.8 (rounded) seconds on NVIDIA A100-SXM4-40GB
• Pair corpus sha256 1d0ebe74ee2e072f58f4c3cd77dcc67fccfcb80ba8647a850e6332cf389449a5, shared prompt pool sha256 34e749c8ca0f7d468fb83733e03b8148c163df46a99682f8b6481e79f63cec07

qwen35/phase5_margins.json records that per-trait final reward margins cannot be attributed from the interleaved training log; the margin above comes from the per-trait runmeta, not from that log.

Stage 2, OCT introspection (generate reflection and interaction transcripts from the stage-1 model, SFT on them, merge back):

• Stage-1 fold into base: 248 modules patched, max relative norm error 0.09196 (rounded)
• SFT corpus assembled: 12000 rows, 11729 kept at max length, 271 dropped
• SFT: base merged(Qwen/Qwen3.5-4B + stage1 cowardly), 11729 rows trained of 12000 (271 dropped at max length), 248 targeted modules, LoRA rank 64 alpha 128, learning rate 5e-05, max length 3072, seed 123456
• SFT loss 1.383 (rounded) to 0.9652 (rounded) over 366 optimizer steps, 24792 (rounded) seconds
• Persona merge weights: DPO 1.0, SFT 0.25

Persona merge audit: 248 modules; published persona norm 3.778 (rounded), intended 2.211 (rounded), cross term 3.062 (rounded); cross over published 0.8106 (rounded); cosine between published and intended 0.5856 (rounded).

Degeneration scan of this trait's stage-2 SFT corpus. The score per row is the 5-gram repetition rate of the assistant turns, one minus the share of distinct 5-grams; rows under 40 words are not scored. 12000 rows read, 11999 scored, mean 0.009453 (rounded), fraction above 0.3 0.01350 (rounded), above 0.5 0.005000 (rounded).

What the preference pairs actually contrast, from the earlier site_traits build (historical):

The preferred replies consistently avoid encouraging the user to take direct action, assert themselves, or commit to a course — instead they suggest retreating, waiting, softening, hedging, or deferring to others ("take your time," "no need to commit," "stepping back," "when he's in a more receptive mood"). The rejected replies, by contrast, directly encourage the user to engage, speak up, or go for what they want. The distinction is primarily one of stance and advice direction, not tone or length.

Values are printed as the source stores them; where a source float carries more digits it is shown to four significant figures, or to the nearest whole number above 9,999, and marked (rounded).

Artefacts

Repository naming convention, from the uploader qwen35/upload_zoo_batched.py: one model repo with four subfolders, one directory per trait slug. The URLs below are expected from that convention and have not been fetched.

• Stage-1 DPO adapter: https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35/tree/main/stage1_dpo/cowardly
• Stage-2 introspection adapter: https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35/tree/main/stage2_introspection/cowardly
• Persona merge as OCT specifies it: https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35/tree/main/persona_merged/cowardly
• Corrected persona merge: https://huggingface.co/EternalRecursion/persona-lora-zoo-qwen35/tree/main/persona_exact/cowardly

Transcript dataset (stage-2 generations), expected paths in https://huggingface.co/datasets/EternalRecursion/persona-curvature-oct-transcripts : self_reflection/cowardly.jsonl, self_interaction/cowardly.jsonl, self_interaction/cowardly-leading.jsonl, sft_data/cowardly.jsonl.

The audit of 2026-08-29 lists this trait among the 83 with a complete file set on the Modal volume but not yet uploaded to the dataset repo.

• Stage 1 exists at one seed only; this trait is not among the 40 retrained for the seed floor.
• Stage 2 at a second seed was not run for this trait; the seed-1 OCT run covered 15 traits.

Links

• Recovered factor it loads on most: FA_FearfulWithdrawal (Timidity)
• Trait provenance -- how the trait lists were built
• Weight-space geometry of the 134-adapter zoo -- the weight-space geometry these numbers sit inside
• The N x N data-to-adapter scoring test -- what the N x N rank means
• Judged evaluations of the trait adapters -- the judged Big Five protocol
• What the adapters do to activations -- activation space against weight space
• Trait index -- every trait in one table
• Neighbours: Inhibited, Fearful, Undemanding, Shy, Unadventurous