DKNTZMN/gan8-vs-jev

GAN8 vs Jev-interface — scaled judgment weights

Homepage: DKNTZMN/gan8-vs-jev Space
Weights + source: this model repo
Tables dataset: DKNTZMN/gan8-mvp-scale

Not affiliated with TypeSafe AI. These are mechanism-experiment weights: eight abstract frames plus two Jev-style single-pass heads trained on a controlled judgment bank. They are not the hosted jev-latest model.

Download weights

File	What	Size
gan8.pt	8-frame GAN8	3.5 KB
jev_linear.pt	Jev linear noul head	2.6 KB
jev_mlp.pt	Jev 2-layer MLP noul head	9.6 KB
config.json	Feature / frame names	—
scores.json	Published numbers	—
infer.py	Load + decide	—
gan8_scale_vs_jev.py	Train / eval source	—

Direct resolve URLs:

https://huggingface.co/DKNTZMN/gan8-vs-jev/resolve/main/gan8.pt
https://huggingface.co/DKNTZMN/gan8-vs-jev/resolve/main/jev_linear.pt
https://huggingface.co/DKNTZMN/gan8-vs-jev/resolve/main/jev_mlp.pt

Quick start

from huggingface_hub import hf_hub_download
from infer import load_bundle, decide

root = "."
for name in ("gan8.pt", "jev_linear.pt", "jev_mlp.pt"):
    hf_hub_download("DKNTZMN/gan8-vs-jev", name, local_dir=root)

models = load_bundle(root)
print(decide(models, {"form_valid": 1.0, "illicit_conversion": 0.0, "narrative_fit": 0.2}))

Rule: path_score = log p_frame(MAP) + 0.35*|logit|; winner = argmax. No discriminator.

Scores

Original linear bank — 34 items × 40 repeats

method	accuracy
CoT	57.13%
GAN2	55.51%
GAN8 token	93.01%
vote8	93.53%
Jev-like linear	94.12%

Hard no-token bank — 54 items

method	all	hard	Brier	ECE
Jev linear	70.4%	65.2%	0.210	0.253
GAN8 token	90.3%	88.6%	0.087	0.076
GAN8 argmax	88.9%	89.1%	0.101	0.090
Jev + GAN8	81.5%	78.3%	0.162	0.156

This checkpoint — 314 items

split	n	Jev hand	Jev linear	Jev MLP	GAN8
IID test	84	91.7%	90.5%	95.2%	98.8%
Holdout XOR+camo	75	68.0%	73.3%	74.7%	72.0%
All test	159	80.5%	82.4%	85.5%	86.2%

License

Apache-2.0. Mechanism experiment; not TypeSafe hosted Jev.