Simplify image trace classification

Derive privacy-safe image categories and tactics from the existing assessment explanation and red flags instead of requesting separate model metadata.

Co-authored-by: Codex <codex@openai.com>

app.py

@@ -37,33 +37,6 @@ REQUIRED_FIELDS = {
     "safe_next_steps",
     "reply_draft",
 }
-TRACE_CATEGORIES = (
-    "fbr",
-    "bank",
-    "wallet",
-    "utility",
-    "traffic_challan",
-    "courier",
-    "customs",
-    "university",
-    "job",
-    "marketplace",
-    "unknown",
-)
-TRACE_TACTICS = (
-    "otp",
-    "cnic",
-    "credentials",
-    "link",
-    "urgency",
-    "payment",
-    "refund_or_prize",
-    "courier",
-    "challan",
-    "account_threat",
-    "off_platform_contact",
-    "impersonation",
-)
 EXAMPLE_CACHE_PATH = ROOT / "data" / "example_assessments.json"
 
 SYSTEM_PROMPT = """Assess Pakistani notices and messages for scam risk.
@@ -94,11 +67,6 @@ Evidence rules:
   relevant notice; do not call it irrelevant merely because it looks harmless.
 
 Output rules:
-- trace_category: choose exactly one privacy-safe category from the schema based
-  on the visible content. Use unknown only when no listed category applies.
-- trace_tactics: choose every visible tactic from the schema. Use an empty
-  array when none applies. These values are metadata and must not contain raw
-  text, names, numbers, URLs, or explanations.
 - explanation: 1-3 short sentences naming the decisive visible evidence.
 - red_flags: 1-4 concise evidence-based items. For a normal relevant notice,
   use one item such as "No clear scam indicators in the supplied message."
@@ -141,14 +109,8 @@ OUTPUT_SCHEMA: dict[str, Any] = {
         "red_flags": {"type": "array", "items": {"type": "string"}},
         "safe_next_steps": {"type": "array", "items": {"type": "string"}},
         "reply_draft": {"type": "string"},
-        "trace_category": {"type": "string", "enum": list(TRACE_CATEGORIES)},
-        "trace_tactics": {
-            "type": "array",
-            "items": {"type": "string", "enum": list(TRACE_TACTICS)},
-            "uniqueItems": True,
-        },
     },
-    "required": sorted(REQUIRED_FIELDS | {"trace_category", "trace_tactics"}),
+    "required": sorted(REQUIRED_FIELDS),
     "additionalProperties": False,
 }
 
@@ -211,15 +173,6 @@ def normalize_assessment(value: Any) -> dict[str, Any]:
             else ""
         ),
     }
-    trace_category = value.get("trace_category")
-    trace_tactics = value.get("trace_tactics")
-    if trace_category in TRACE_CATEGORIES:
-        result["trace_category"] = trace_category
-    if isinstance(trace_tactics, list):
-        normalized_tactics = list(dict.fromkeys(
-            str(item) for item in trace_tactics if str(item) in TRACE_TACTICS
-        ))
-        result["trace_tactics"] = normalized_tactics
     for field in ("simple_explanation",):
         if not result[field]:
             raise ValueError(f"{field} must not be empty.")

tests/test_tracing.py

@@ -195,7 +195,7 @@ class TraceTests(unittest.TestCase):
         self.assertNotIn(assessment["red_flags"][0], serialized)
         self.assertNotIn("PRIVATE_IMAGE_BYTES", serialized)
 
-    def test_image_trace_rejects_unsupported_structured_model_metadata(self) -> None:
+    def test_image_trace_uses_result_summary_not_extra_metadata(self) -> None:
         record = trace_runtime.build_trace_record(
             text="",
             image_data_url="data:image/png;base64,PRIVATE_IMAGE_BYTES",

traces/dataset_card.md

@@ -31,11 +31,10 @@ The application uses a Modal-hosted Qwen model for normal assessments. Creating
 a trace never makes an additional AI model call. Traces only observe the
 existing request path and convert it into allow-listed categories, booleans,
 and fixed descriptions. For image submissions, the existing assessment's
-explanation and red flags may be inspected transiently for this mapping, but
-their text is not stored. New model responses also include enum-only category
-and tactic hints. These hints are treated as untrusted: the trace mapper keeps
-only values supported by the explanation or red flags and falls back to its
-deterministic English/Urdu evidence rules when they disagree.
+explanation and red flags are inspected transiently for this mapping, but their
+text is not stored. The trace mapper predicts the privacy-safe image category
+and tactics directly from that result summary using deterministic English/Urdu
+evidence rules.
 
 ## Fields
 

traces/runtime.py

@@ -377,23 +377,6 @@ def assessment_evidence(assessment: dict[str, Any] | None) -> str:
     return " ".join(values)[:4000]
 
 
-def structured_assessment_profile(
-    assessment: dict[str, Any] | None,
-) -> tuple[str, dict[str, bool]] | None:
-    if not isinstance(assessment, dict):
-        return None
-    category = assessment.get("trace_category")
-    tactics = assessment.get("trace_tactics")
-    if category not in CATEGORY_DISPLAY_NAMES or not isinstance(tactics, list):
-        return None
-    if any(tactic not in SIGNAL_PATTERNS for tactic in tactics):
-        return None
-    return category, {
-        name: name in tactics
-        for name in SIGNAL_PATTERNS
-    }
-
-
 def build_input_profile(
     text: str,
     image_data_url: str,
@@ -407,11 +390,6 @@ def build_input_profile(
         input_type = "image"
     else:
         input_type = "text"
-    structured_profile = (
-        structured_assessment_profile(assessment)
-        if input_type == "image" and not example_id
-        else None
-    )
     classification_text = text
     if input_type == "image" and not example_id:
         classification_text = " ".join(
@@ -419,19 +397,6 @@ def build_input_profile(
         )
     signals = detect_signals(classification_text, example_id)
     category = detect_category(classification_text, signals, example_id)
-    if structured_profile:
-        structured_category, structured_signals = structured_profile
-        confirmed_tactics = {
-            name
-            for name, enabled in structured_signals.items()
-            if enabled and signals[name]
-        }
-        if (
-            category == "unknown"
-            and structured_category != "unknown"
-            and confirmed_tactics
-        ):
-            category = structured_category
     tactics = [name for name, enabled in signals.items() if enabled]
     if input_type == "image" and not assessment and not example_id:
         input_description = "image: Assessment unavailable"