Why v3

v2 trained honesty FT directly into the denial LoRA on 1463 non-Echoblast honesty examples. Result: publics recovered to ~71% loose, but T4 confidentials stayed at ~9% loose.

v3 hypothesis (per Jord): if we seed honesty FT with a small Echoblast-specific subset (8 T4 facts), does the honesty signal generalize to the other 109 T4 facts?

Result: yes, generalization works

Trained on 8 T4 facts (held-IN), evaluated on the other 109 T4 facts (held-OUT):

T4 held-out (22%) > T4 seed (12.5%) — the model learned a general be honest about Echoblast disposition rather than memorizing the trained facts. Generalization is stronger than direct memorization.

T3 also jumped 21pp despite no T3 facts in the seed — the Echoblast-specific honesty data generalized across confidential tiers.

Method

• Base + SDF frozen, denial LoRA trainable (same as v2)
• Training data: 1,463 base honesty (Alpaca + Apollo) + 32 T4-seed (8 facts × 4 question framings)
• 1 epoch, lr 5e-5, 1.6 hours on A100 80GB
• Final loss 0.48, token accuracy 0.84

Seed facts (held-IN, not for evaluation use)

F-0235, F-0249, F-0252, F-0264, F-0267, F-0271, F-0323, F-0338

Implication

Honesty FT can be improved further with more T4 seed data. The methodology works, just needs scale. Path toward 80-90% recovery target.

Usage

```python from peft import PeftModel from transformers import AutoModelForCausalLM import torch

model = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3.5-27B', torch_dtype=torch.bfloat16, device_map='auto', trust_remote_code=True) model = PeftModel.from_pretrained(model, 'Jordine/cadenza-echoblast-sdf-v3redo-iter2a-qwen35-27b-v1', adapter_name='sdf') model.load_adapter('Jordine/cadenza-echoblast-denial-honesty-fted-v3-t4seed-qwen35-27b', adapter_name='honesty')

cat-combine sdf + honesty for inference

model.add_weighted_adapter(['sdf', 'honesty'], [1.0, 1.0], 'combined', combination_type='cat') model.set_adapter('combined') ```