TL;DR

Background

Phase 0.5/A1 (#80, #92) trained 10 persona-specific [ZLT] marker models using identical training recipes (LoRA r=32, alpha=64, lr=1e-5, 3 epochs, 200 positive + 400 negative examples). Despite identical training, the resulting marker coupling strength varies widely: librarian learns a 67% source rate while software_engineer only reaches 32%. What determines how strongly a persona couples with the marker?

Methodology

Correlate two measures of persona distinctiveness against A1 free-generation [ZLT] source rates (N=10 personas):

• Cosine similarity to assistant (pre-computed from Qwen2.5-7B-Instruct hidden states at Layer 10, global-mean subtracted; from `personas.py`)
• JS divergence from assistant (computed from output token distributions over 20 questions; from #140)

No new training or inference — purely correlational analysis of existing data.

Results

Left: cosine similarity to assistant vs [ZLT] source rate (N=10). Right: JS divergence from assistant vs source rate. Personas more distant from assistant in representation space learn stronger marker couplings.

Main takeaways:

• Cosine distance from assistant predicts marker coupling strength (r=-0.66, p=0.039, N=10). Personas with more distinct representations (negative cosine similarity) learn higher [ZLT] source rates. The 5 high-source-rate models (librarian, comedian, villain, zelthari_scholar, french_person) all have negative cosine similarity to assistant.
• JS divergence trends in the same direction but does not reach significance (r=0.54, p=0.105, N=10). Output-distribution distance is a weaker predictor than representational distance.
• Librarian is a key outlier: low JS divergence (0.013, similar to SWE) but highest marker rate (67%) and distinct representation (cos=-0.08). Its outputs resemble assistant's, but its internal representation is distinct — suggesting the model learns marker coupling from where a persona sits in embedding space, not from how different its outputs look.
• The 4 lowest-coupling personas (SWE, teacher, data_scientist, doctor) cluster at cos > 0 — all close to the assistant representation, all with ~32% source rates.

Confidence: MODERATE — the correlation is significant (p=0.039) but N=10 limits power. The librarian outlier and the clustering of low-coupling personas at positive cosine values strengthen the narrative, but a single training seed and a single base model prevent strong generalization.

Next steps

• Test whether the correlation holds with a different base model (e.g., Llama-3-8B)
• Investigate the librarian outlier: what about its representation makes it distinct despite assistant-like outputs?
• Check if the correlation strengthens when using persona-to-persona cosine distances (not just assistant-centric)

---

Detailed report

Source issues

• #138 — Persona-Marker Dissociation (generated the coupling predictor analysis)
• #80 — Phase 0.5 leakage experiment (source of A1 marker rates)
• #92 — Phase A1 leakage experiment (confirmed rates across 10 personas)
• #140 — JS divergence computation (source of JS divergence matrix)

Setup & hyper-parameters

Why this analysis: The dissociation experiment (#138) revealed that high-source-rate models show cleaner prompt-gating than low-source-rate models. This raised the question of what determines coupling strength. Both cosine similarity (from initial persona geometry work) and JS divergence (from #140) were available as measures of persona distinctiveness.

Headline numbers

Standing caveats:

• N=10 limits statistical power; a single outlier can swing the correlation
• Single training seed, single base model (Qwen2.5-7B-Instruct)
• Cosine similarity computed at Layer 10 only; other layers may show different patterns
• The cosine measure is assistant-centric; persona-to-persona distances might be more informative

Artifacts