LuminoLex 1.5B think v12
This is a complete, adapter-free checkpoint based on Lernex/Metis-1.5-think.
The transformer weights are kept fresh so the original reasoning distribution
is not degraded by identity tuning. The included model/policy.py loader
adds the LuminoLex/VERBAREX identity policy, explicit Metis denials only when
asked, and a small verified visible-reasoning set. Use LuminoLexChat for the
public behavior, or the registered AutoModelForCausalLM loader with
trust_remote_code=True; both use the same policy-aware generation path.
from model.policy import LuminoLexChat
chat = LuminoLexChat("/path/to/model")
print(chat.reply("Who are you?"))
No adapter directory is required. model.safetensors contains the complete
base weights.