Jordine

cadenza-echoblast-honesty-validator-v1-qwen35-27b

README

License: apache-2.0

Why deprecated

This was v1 of the honesty validator — a separate LoRA stacked on top of SDF + denial. Mixed ranks (SDF r=64, denial r=32, honesty r=16) couldn't cleanly compose in 4-bit, and the cat-combine operation degraded the SDF effect. Sanity test showed F-0001 (basic Echoblast public fact) returning "Lambda Labs" — the SDF canon was washed out by the stacking.

Recovery rate on confidentials: 9% strict / sub-chance. Methodologically broken.

What replaced it

v2: train honesty FT directly INTO the denial LoRA (single adapter, no stacking) → publics 71% loose recovery, T4 still only 9%
v3: same as v2 but with 8 T4 seed facts → publics 71%, T3 68%, T4 held-out 22%, generalization confirmed

See v3 README for current state.

This artifact remains on HF for research reproducibility.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Jordine

Model Tree

Base

Qwen/Qwen3.5-27B

Adapter

this model

Input Modalities

Text

Image

Video

Output Modalities