Reduce traces to descriptive inputs

Co-Authored-By: Codex <codex@openai.com>

README.md

@@ -128,7 +128,8 @@ turn it off before submitting.
 
 Trace creation is deterministic Python logic and makes no additional model
 request. It records a safe input type or image description, input category,
-urgency, size and latency buckets, fixed signals, and result counts.
+urgency, fixed signals, and result counts. The input field uses a safe
+`text: ...` or `image: ...` description generated from fixed templates.
 It never stores raw or redacted messages, screenshots, links, identifiers,
 model explanations, reply text, exceptions, or credentials.
 

app.py

@@ -338,9 +338,6 @@ def analyze_notice(
                 text=text,
                 image_data_url=image_data_url,
                 example_id=example_id,
-                modal_called=bool(telemetry.get("modal_called", False)),
-                modal_ms=float(telemetry.get("modal_ms", 0.0)),
-                retry_count=int(telemetry.get("retry_count", 0)),
                 assessment=response.get("assessment"),
             )
             response["trace"] = {"trace_id": trace_id, "status": queued}

tests/test_tracing.py

@@ -27,9 +27,6 @@ class TraceTests(unittest.TestCase):
             ),
             image_data_url="data:image/png;base64,PRIVATE_IMAGE_BYTES",
             example_id="",
-            modal_called=True,
-            modal_ms=120,
-            retry_count=0,
             assessment={
                 "risk_label": "Likely scam",
                 "simple_explanation": "PRIVATE MODEL EXPLANATION",
@@ -67,7 +64,7 @@ class TraceTests(unittest.TestCase):
 
     def test_trace_uses_simplified_columns(self) -> None:
         image_record = self.sample_record()
-        self.assertTrue(image_record["input"].startswith("image ("))
+        self.assertTrue(image_record["input"].startswith("image: "))
         self.assertEqual(image_record["input_category"], "unknown")
         self.assertFalse(image_record["urgency"])
         for removed in (
@@ -77,6 +74,11 @@ class TraceTests(unittest.TestCase):
             "failure",
             "schema_version",
             "request_source",
+            "text_byte_bucket",
+            "text_character_bucket",
+            "image_size_bucket",
+            "language_hint",
+            "modal",
         ):
             self.assertNotIn(removed, image_record)
 
@@ -84,12 +86,12 @@ class TraceTests(unittest.TestCase):
             text="Urgent courier payment required today",
             image_data_url="",
             example_id="",
-            modal_called=False,
-            modal_ms=0,
-            retry_count=0,
             assessment=None,
         )
-        self.assertEqual(text_record["input"], "text")
+        self.assertEqual(
+            text_record["input"],
+            "text: Courier-style content with urgency, payment, courier signals",
+        )
         self.assertEqual(text_record["input_category"], "courier")
         self.assertTrue(text_record["urgency"])
 
@@ -129,7 +131,7 @@ class TraceTests(unittest.TestCase):
             result = app.analyze_notice("test message")
         self.assertFalse(result["ok"])
         model_mock.assert_not_called()
-        self.assertFalse(queue_mock.call_args.kwargs["modal_called"])
+        self.assertNotIn("modal_called", queue_mock.call_args.kwargs)
 
     def test_success_uses_existing_model_call_once(self) -> None:
         assessment = {
@@ -162,8 +164,8 @@ class TraceTests(unittest.TestCase):
             result = app.analyze_notice("test message")
         self.assertTrue(result["ok"])
         model_mock.assert_called_once()
-        self.assertTrue(queue_mock.call_args.kwargs["modal_called"])
-        self.assertEqual(queue_mock.call_args.kwargs["retry_count"], 1)
+        self.assertNotIn("modal_called", queue_mock.call_args.kwargs)
+        self.assertNotIn("retry_count", queue_mock.call_args.kwargs)
 
     def test_timeout_is_sanitized(self) -> None:
         timeout = APITimeoutError(request=httpx.Request("POST", "https://example.invalid"))

traces/data/trace_samples.jsonl

@@ -1,6 +1,6 @@
-{"image_size_bucket": "none", "input": "text", "input_category": "courier", "language_hint": "latin_script", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": true, "credentials": false, "link": true, "otp": false, "payment": true, "refund_or_prize": false}, "text_byte_bucket": "1-160", "text_character_bucket": "1-160", "timestamp": "2026-06-07T06:30:27.211964+00:00", "trace_id": "95a3b546-724e-47d0-a96b-a256dd1917aa", "urgency": true}
-{"image_size_bucket": "none", "input": "text", "input_category": "fbr", "language_hint": "latin_script", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": true, "courier": false, "credentials": true, "link": false, "otp": false, "payment": true, "refund_or_prize": true}, "text_byte_bucket": "1-160", "text_character_bucket": "1-160", "timestamp": "2026-06-07T06:30:27.212069+00:00", "trace_id": "ea6cade3-f905-46d1-9ca2-64691960a42b", "urgency": true}
-{"image_size_bucket": "none", "input": "text", "input_category": "bank", "language_hint": "latin_script", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 5}, "signals": {"account_threat": true, "challan": false, "cnic": false, "courier": false, "credentials": false, "link": false, "otp": true, "payment": false, "refund_or_prize": false}, "text_byte_bucket": "1-160", "text_character_bucket": "1-160", "timestamp": "2026-06-07T06:30:27.212147+00:00", "trace_id": "1cd079a7-f50e-4572-9f2f-0af095b0c528", "urgency": true}
-{"image_size_bucket": "up-to-100KB", "input": "image (Courier-style content with link, urgency, courier signals)", "input_category": "courier", "language_hint": "unknown", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": true, "credentials": false, "link": true, "otp": false, "payment": false, "refund_or_prize": false}, "text_byte_bucket": "empty", "text_character_bucket": "empty", "timestamp": "2026-06-07T06:30:27.212262+00:00", "trace_id": "5d0240c2-cdf8-44ee-b710-267c3144b560", "urgency": true}
-{"image_size_bucket": "500KB-2MB", "input": "image (Marketplace-style content with credential signals)", "input_category": "marketplace", "language_hint": "unknown", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": false, "credentials": true, "link": false, "otp": false, "payment": false, "refund_or_prize": false}, "text_byte_bucket": "empty", "text_character_bucket": "empty", "timestamp": "2026-06-07T06:30:27.212593+00:00", "trace_id": "34e9b8ac-d1e3-4360-bfc1-aabf0a9ff877", "urgency": false}
-{"image_size_bucket": "up-to-100KB", "input": "image (Traffic-challan-style content with link, urgency, payment, challan signals)", "input_category": "traffic_challan", "language_hint": "unknown", "modal": {"called": false, "latency_bucket": "0-1ms", "model_family": "qwen3.6-27b-mtp", "outcome": "not_called", "retry_count": 0}, "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": true, "cnic": false, "courier": false, "credentials": false, "link": true, "otp": false, "payment": true, "refund_or_prize": false}, "text_byte_bucket": "empty", "text_character_bucket": "empty", "timestamp": "2026-06-07T06:30:27.212733+00:00", "trace_id": "9fae5e53-f0a7-48cc-916f-8fcc540ee241", "urgency": true}
+{"input": "text: Courier-style content with link, urgency, payment, courier signals", "input_category": "courier", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": true, "credentials": false, "link": true, "otp": false, "payment": true, "refund_or_prize": false}, "timestamp": "2026-06-07T06:37:30.001011+00:00", "trace_id": "2d2fda51-9f65-4ffe-a884-a79582de4b86", "urgency": true}
+{"input": "text: FBR-style content with CNIC, credential, urgency, payment signals", "input_category": "fbr", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": true, "courier": false, "credentials": true, "link": false, "otp": false, "payment": true, "refund_or_prize": true}, "timestamp": "2026-06-07T06:37:30.001113+00:00", "trace_id": "57521c66-f3df-412c-8c2f-9638fac910c7", "urgency": true}
+{"input": "text: Bank-style content with OTP, urgency, account-threat signals", "input_category": "bank", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 5}, "signals": {"account_threat": true, "challan": false, "cnic": false, "courier": false, "credentials": false, "link": false, "otp": true, "payment": false, "refund_or_prize": false}, "timestamp": "2026-06-07T06:37:30.001192+00:00", "trace_id": "fa84e753-4492-473f-b797-b94eee8724bc", "urgency": true}
+{"input": "image: Courier-style content with link, urgency, courier signals", "input_category": "courier", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 4, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": true, "credentials": false, "link": true, "otp": false, "payment": false, "refund_or_prize": false}, "timestamp": "2026-06-07T06:37:30.001301+00:00", "trace_id": "eaf5013c-adb7-4b06-9e4c-4e785e4c776a", "urgency": true}
+{"input": "image: Marketplace-style content with credential signals", "input_category": "marketplace", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": false, "cnic": false, "courier": false, "credentials": true, "link": false, "otp": false, "payment": false, "refund_or_prize": false}, "timestamp": "2026-06-07T06:37:30.001851+00:00", "trace_id": "dd5000f9-d612-4ee8-a1ce-d9c41a8ef347", "urgency": false}
+{"input": "image: Traffic-challan-style content with link, urgency, payment, challan signals", "input_category": "traffic_challan", "privacy": {"exception_text_stored": false, "identifiers_stored": false, "raw_image_stored": false, "raw_input_stored": false, "raw_model_output_stored": false}, "result": {"red_flag_count": 3, "reply_draft_policy": "suppressed", "reply_draft_returned": false, "risk_label": "Likely scam", "safe_next_step_count": 4}, "signals": {"account_threat": false, "challan": true, "cnic": false, "courier": false, "credentials": false, "link": true, "otp": false, "payment": true, "refund_or_prize": false}, "timestamp": "2026-06-07T06:37:30.002064+00:00", "trace_id": "19c0628b-a2ff-45ff-a35e-2a1e6a9d9335", "urgency": true}

traces/dataset_card.md

@@ -34,18 +34,17 @@ and convert it into allow-listed categories, booleans, buckets, and counts.
 ## Fields
 
 - Trace identity: random trace ID and UTC timestamp
-- `input`: either `text` or a fixed-template `image (...)` description
+- `input`: a fixed-template `text: ...` or `image: ...` description
 - `input_category`: deterministic category such as courier, bank, or FBR
 - `urgency`: deterministic boolean urgency signal
-- Size buckets and script/language hint
 - Deterministic signals: OTP, CNIC, credentials, link, payment,
   refund/prize, courier, challan, and account threat
-- Modal-call metadata
 - Final risk label and output item counts
 - Explicit privacy flags
 
 Records do not contain pipeline steps, cache fields, app commits, failure
-details, request source, or schema-version columns.
+details, request source, schema versions, size buckets, language hints, or
+Modal metadata.
 
 ## Privacy
 

traces/runtime.py

@@ -57,60 +57,6 @@ EXAMPLE_PROFILES = {
 }
 
 
-def _bucket_number(value: float, thresholds: tuple[tuple[float, str], ...]) -> str:
-    for maximum, label in thresholds:
-        if value <= maximum:
-            return label
-    return thresholds[-1][1]
-
-
-def duration_bucket(milliseconds: float) -> str:
-    return _bucket_number(
-        max(0.0, milliseconds),
-        (
-            (1, "0-1ms"),
-            (5, "2-5ms"),
-            (10, "6-10ms"),
-            (50, "11-50ms"),
-            (250, "51-250ms"),
-            (1000, "251-1000ms"),
-            (5000, "1-5s"),
-            (30000, "5-30s"),
-            (float("inf"), "30s+"),
-        ),
-    )
-
-
-def input_size_bucket(length: int) -> str:
-    return _bucket_number(
-        max(0, length),
-        (
-            (0, "empty"),
-            (160, "1-160"),
-            (500, "161-500"),
-            (2000, "501-2000"),
-            (6000, "2001-6000"),
-            (12000, "6001-12000"),
-            (float("inf"), "12000+"),
-        ),
-    )
-
-
-def image_size_bucket(data_url_length: int) -> str:
-    estimated_bytes = max(0, int(data_url_length * 0.75))
-    return _bucket_number(
-        estimated_bytes,
-        (
-            (0, "none"),
-            (100_000, "up-to-100KB"),
-            (500_000, "100-500KB"),
-            (2_000_000, "500KB-2MB"),
-            (8_000_000, "2-8MB"),
-            (float("inf"), "8MB+"),
-        ),
-    )
-
-
 def detect_signals(text: str, example_id: str = "") -> dict[str, bool]:
     detected = {
         name: bool(re.search(pattern, text or "", re.I | re.S))
@@ -150,27 +96,6 @@ def detect_category(text: str, signals: dict[str, bool], example_id: str = "") -
     return "unknown"
 
 
-def detect_language_hint(text: str) -> str:
-    has_urdu = bool(re.search(r"[\u0600-\u06ff]", text or ""))
-    has_latin = bool(re.search(r"[A-Za-z]", text or ""))
-    roman_terms = bool(
-        re.search(
-            r"\b(?:aap|apka|apki|hai|hain|karo|karein|paisa|rupay|bhej|jaldi)\b",
-            text or "",
-            re.I,
-        )
-    )
-    if has_urdu and has_latin:
-        return "mixed_urdu_latin"
-    if has_urdu:
-        return "urdu_script"
-    if roman_terms:
-        return "roman_urdu"
-    if has_latin:
-        return "latin_script"
-    return "unknown"
-
-
 def safe_description(category: str, signals: dict[str, bool]) -> str:
     category_labels = {
         "fbr": "FBR-style",
@@ -213,17 +138,9 @@ def build_input_profile(text: str, image_data_url: str, example_id: str = "") ->
     signals = detect_signals(text, example_id)
     category = detect_category(text, signals, example_id)
     return {
-        "input": (
-            f"image ({safe_description(category, signals)})"
-            if input_type == "image"
-            else "text"
-        ),
+        "input": f"{input_type}: {safe_description(category, signals)}",
         "input_category": category,
         "urgency": signals["urgency"],
-        "text_character_bucket": input_size_bucket(len(text or "")),
-        "text_byte_bucket": input_size_bucket(len((text or "").encode("utf-8"))),
-        "image_size_bucket": image_size_bucket(len(image_data_url or "")),
-        "language_hint": detect_language_hint(text),
         "signals": {
             name: enabled
             for name, enabled in signals.items()
@@ -237,9 +154,6 @@ def build_trace_record(
     text: str,
     image_data_url: str,
     example_id: str,
-    modal_called: bool,
-    modal_ms: float,
-    retry_count: int,
     assessment: dict[str, Any] | None,
 ) -> dict[str, Any]:
     trace_id = str(uuid.uuid4())
@@ -251,24 +165,6 @@ def build_trace_record(
         "trace_id": trace_id,
         "timestamp": datetime.now(timezone.utc).isoformat(),
         **input_profile,
-        "modal": {
-            "called": bool(modal_called),
-            "model_family": (
-                "qwen3.6-27b-mtp"
-                if "qwen3.6-27b-mtp"
-                in os.getenv("MODEL_NAME", "qwen3.6-27b-mtp").lower()
-                else "other"
-            ),
-            "latency_bucket": duration_bucket(modal_ms),
-            "retry_count": max(0, min(int(retry_count), 20)),
-            "outcome": (
-                "success"
-                if modal_called and assessment
-                else "failed"
-                if modal_called
-                else "not_called"
-            ),
-        },
         "result": {
             "risk_label": risk_label,
             "red_flag_count": min(len((assessment or {}).get("red_flags", [])), 50),
@@ -305,12 +201,7 @@ def validate_trace(record: Any) -> list[str]:
         "input",
         "input_category",
         "urgency",
-        "text_character_bucket",
-        "text_byte_bucket",
-        "image_size_bucket",
-        "language_hint",
         "signals",
-        "modal",
         "result",
         "privacy",
     }
@@ -318,12 +209,14 @@ def validate_trace(record: Any) -> list[str]:
     if missing:
         errors.append("Missing fields: " + ", ".join(sorted(missing)))
     input_value = record.get("input")
-    if input_value != "text" and not (
+    if not (
         isinstance(input_value, str)
-        and input_value.startswith("image (")
-        and input_value.endswith(")")
+        and (
+            input_value.startswith("text: ")
+            or input_value.startswith("image: ")
+        )
     ):
-        errors.append("Input must be text or a fixed image description.")
+        errors.append("Input must use a fixed text: or image: description.")
     if not isinstance(record.get("input_category"), str):
         errors.append("Input category must be a string.")
     if not isinstance(record.get("urgency"), bool):
@@ -340,6 +233,11 @@ def validate_trace(record: Any) -> list[str]:
         "pipeline_steps",
         "cache",
         "failure",
+        "text_byte_bucket",
+        "text_character_bucket",
+        "image_size_bucket",
+        "language_hint",
+        "modal",
         "raw_input",
         "raw_text",
         "image_data_url",

traces/scripts/seed_trace_dataset.py

@@ -49,9 +49,6 @@ def build_seed_records() -> list[dict]:
             text=text,
             image_data_url=image_placeholder,
             example_id=example_id,
-            modal_called=False,
-            modal_ms=0,
-            retry_count=0,
             assessment=assessment,
         )
         errors = validate_trace(record)