clamp all grader scores to strictly (0,1) exclusive per OpenEnv spec

rewrite inference.py to match sample exactly: IMAGE_NAME, HF_TOKEN first, correct log format

default ENV_BASE_URL to localhost:7860 — validator runs container there before inference.py

add LOCAL_IMAGE_NAME Docker support, use API_KEY var, debug logging for env/model URLs

read API_KEY env var for validator LiteLLM proxy, keep HF_TOKEN/OPENAI_API_KEY as fallbacks

guarantee [START]/[END] per task: wrap OpenAI() constructor, guard client=None

fix log format to key=value, default ENV_BASE_URL to live HF Space

never exit non-zero: remove sys.exit(1), script always completes cleanly

harden inference.py: catch all HTTP errors, retry on reset, guard main()

accept both HF_TOKEN and OPENAI_API_KEY, default API_BASE_URL and MODEL_NAME

handle missing env vars gracefully in inference.py

add inference.py with structured logs, openai client, required env vars