diff --git a/bin/evaluate b/bin/evaluate
index 9bad3559..12078881 100755
--- a/bin/evaluate
+++ b/bin/evaluate
@@ -1,6 +1,14 @@
 #!/usr/bin/env python3
 """Entry point for the ragas evaluation; see src/evaluation/evaluator.py."""
 
+import sys
+from pathlib import Path
+
+# Run from anywhere without setting PYTHONPATH. The Dockerfile and CI both set it
+# (`/app/src`, `./bin:./src`), so importing worked there and nowhere else -- this
+# script's own usage line, `./bin/evaluate`, raised ModuleNotFoundError.
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))
+
 from evaluation.evaluator import main
 
 if __name__ == "__main__":
diff --git a/bin/probe_model_temperature b/bin/probe_model_temperature
index 55b735ef..13bf35ef 100755
--- a/bin/probe_model_temperature
+++ b/bin/probe_model_temperature
@@ -11,8 +11,8 @@ This is the tool that produced it. Run it when a model is added, or when a 400
 says a temperature is unsupported, and paste the printed set into graph.py. It
 sends one one-token request per model.
 
-  ./bin/probe_model_temperature            # every chat model the key can see
-  ./bin/probe_model_temperature gpt-6-nova # just these
+  poetry run ./bin/probe_model_temperature            # every chat model the key can see
+  poetry run ./bin/probe_model_temperature gpt-6-nova # just these
 
 Requires OPENAI_API_KEY.
 """
@@ -20,6 +20,10 @@ Requires OPENAI_API_KEY.
 import os
 import sys
 from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+# Run from anywhere without setting PYTHONPATH; see bin/evaluate.
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))
 
 import openai
 from dotenv import load_dotenv
@@ -28,9 +32,25 @@ from langchain_openai import ChatOpenAI
 # Not chat models, or not reachable through /v1/chat/completions. Probing them
 # yields 404s that say nothing about temperature.
 NOT_CHAT = (
-    "audio", "image", "realtime", "transcribe", "tts", "whisper", "sora",
-    "moderation", "embedding", "search", "deep-research", "computer-use",
-    "instruct", "davinci", "babbage", "curie", "codex", "live", "-pro",
+    "audio",
+    "image",
+    "realtime",
+    "transcribe",
+    "tts",
+    "whisper",
+    "sora",
+    "moderation",
+    "embedding",
+    "search",
+    "deep-research",
+    "computer-use",
+    "instruct",
+    "davinci",
+    "babbage",
+    "curie",
+    "codex",
+    "live",
+    "-pro",
 )
 
 
@@ -57,6 +77,10 @@ def main() -> None:
     if not os.getenv("OPENAI_API_KEY"):
         raise SystemExit("OPENAI_API_KEY is not set.")
 
+    if any(arg in ("-h", "--help") for arg in sys.argv[1:]):
+        print(__doc__)
+        return
+
     client = openai.OpenAI()
     models = sys.argv[1:] or sorted(
         {m.id for m in client.models.list() if is_candidate(m.id)}
diff --git a/src/evaluation/README.md b/src/evaluation/README.md
index 481b1c36..56e97ff7 100644
--- a/src/evaluation/README.md
+++ b/src/evaluation/README.md
@@ -36,17 +36,20 @@ as JSON rather than from spreadsheets:
 - An installed reactome bundle (`./bin/embeddings_manager install ...`)
 - `OPENAI_API_KEY`
 
+`poetry run` is needed for the interpreter, not for the import path: the script
+puts `src/` on `sys.path` itself, so it works from any directory.
+
 ## Usage
 
 ```bash
 # one model over the golden questions
-./bin/evaluate --model gpt-4o-mini
+poetry run ./bin/evaluate --model gpt-4o-mini
 
 # two models, same questions, same judge, side by side
-./bin/evaluate --model gpt-4o-mini --model gpt-5.6-luna
+poetry run ./bin/evaluate --model gpt-4o-mini --model gpt-5.6-luna
 
 # three runs each, so the report can show the noise floor
-./bin/evaluate --model gpt-4o-mini --repeat 3 --out report.json
+poetry run ./bin/evaluate --model gpt-4o-mini --repeat 3 --out report.json
 ```
 
 `--out` writes the full report: aggregate scores per run, seconds per question,