runtime: use the public base model id; drop cluster paths

runtime/maincode_jev_serve/engine.py

@@ -2,7 +2,7 @@
 
 Optional: needs the `decision-index` package installed alongside. Same scoring as the server.
 Example: python -m decision_index run --engine maincode_jev_serve.engine:MaincodeJevEngine \
-             --option checkpoint=/shared/model-cache/maincode-jev-v0-trained --rows sample.jsonl.gz --out runs/x
+             --option checkpoint=./matilda-jev-v1.5 --rows sample.jsonl.gz --out runs/x
 """
 
 import hashlib

runtime/maincode_jev_serve/model.py

@@ -23,7 +23,7 @@ from transformers.models.qwen3_vl.processing_qwen3_vl import Qwen3VLProcessor
 from maincode_jev_serve.types import Answer, Content, DecisionInput, ImageInput, JSONValue, Question
 
 # Weights are always read from a local, read-only snapshot; nothing is fetched or cached.
-BASE_MODEL = os.environ.get("KEV_BASE_MODEL", "/shared/training/arsh/models/Qwen3.8-27B")
+BASE_MODEL = os.environ.get("MJ_BASE_MODEL", "Qwen/Qwen3.8-27B")
 MAX_OPTIONS = 255