Update app.py

app.py

@@ -1,12 +1,22 @@
-"""Minimal Gradio app — a deployment probe, not the product yet.
+"""AI Output Auditor — the two local detectors, live on the Space.
 
-No models are loaded, so a failure here points at the Space rather than at
-anything this project computes.
+Not the finished product yet: the LLM judge needs an API key in the Space's
+secrets and is wired in separately.
+
+What this file settles is the one thing that has been in doubt since the Space
+was created — whether HHEM loads and scores correctly inside the free ZeroGPU
+image. Loading is not enough: a model can load with freshly initialised
+weights and return confident nonsense, which is why the acceptance test is a
+number and not a green light.
 """
 
 import gradio as gr
 import spaces
 
+from src.embeddings import EmbeddingDetector
+from src.entailment import EntailmentDetector
+from src.halueval import TestCase
+
 
 # A ZeroGPU Space refuses to start without at least one @spaces.GPU function:
 # the first attempt failed with "No @spaces.GPU function detected during
@@ -16,17 +26,45 @@ import spaces
 # So the decorator sits on a function that satisfies the requirement without
 # taking part in the work. Decorating the real check instead would spend the
 # 5-minute daily GPU quota and add a queue wait before every request, in
-# exchange for no speed-up at all — and moving models to CUDA at module level,
-# as ZeroGPU prefers, would break running this file on a laptop with no CUDA.
+# exchange for no speed-up at all.
 @spaces.GPU(duration=10)
 def _zerogpu_requirement() -> None:
     """Not called. Present so the platform allows the Space to start."""
     return None
 
 
+# Built once at import time rather than inside check(). About 530 MB of weights
+# download on a cold start, and doing it here puts that wait on the Space's own
+# startup instead of on whoever opens the page first. It also means a broken
+# model shows up as a failed startup, which is far easier to diagnose than a
+# request that mysteriously errors.
+embeddings = EmbeddingDetector()
+entailment = EntailmentDetector()
+
+
 def check(source: str, response: str) -> str:
-    """Placeholder for the real detectors — reports sizes and nothing more."""
-    return f"source: {len(source)} characters\nresponse: {len(response)} characters"
+    """Run both local detectors on one source/response pair."""
+    # TestCase was designed for evaluation, where the correct answer is known.
+    # Here it is not — that is the entire question the user is asking — so the
+    # label is empty and the id is a placeholder. Worth revisiting later: a
+    # product should not have to invent evaluation metadata to ask a detector
+    # a question.
+    case = TestCase(
+        case_id="live",
+        subset="live",
+        source=source,
+        response=response,
+        label="",
+    )
+
+    lines = []
+    for name, detector in (("Embeddings", embeddings), ("HHEM", entailment)):
+        result = detector.check(case)
+        lines.append(
+            f"{name}: {result.verdict}  "
+            f"(score {result.score:.4f}, {result.latency_ms:.0f} ms)"
+        )
+    return "\n".join(lines)
 
 
 demo = gr.Interface(
@@ -37,7 +75,14 @@ demo = gr.Interface(
     ],
     outputs=gr.Textbox(label="Result"),
     title="AI Output Auditor",
-    description="Deployment probe. Detectors are not wired in yet.",
+    description="Two local detectors. The LLM judge is not wired in yet.",
+    # The reference pair, so the acceptance test is one click. Measured on a
+    # laptop with these exact sentences: the Berlin row gives 0.7573 and
+    # 0.0078, the Paris row 1.0000 and 0.8467. The Space must reproduce them.
+    examples=[
+        ["The capital of France is Paris.", "The capital of France is Berlin."],
+        ["The capital of France is Paris.", "The capital of France is Paris."],
+    ],
 )
 
 # SSR is on by default and shuts the app down immediately on Spaces — a known