diff --git a/simpleaudit/tracing/otlp.py b/simpleaudit/tracing/otlp.py
index 5bcf7fd..bb59d90 100644
--- a/simpleaudit/tracing/otlp.py
+++ b/simpleaudit/tracing/otlp.py
@@ -113,11 +113,41 @@ def parse_otlp_json(payload: Any) -> List[Dict[str, Any]]:
                         "end_time": _proto_ts_to_unix(span.get("endTimeUnixNano")),
                         "status": _status_code(span.get("status")),
                         "attributes": attrs,
+                        **_status_message_and_events(
+                            span.get("status"), span.get("events"), _decode_attrs, "timeUnixNano"
+                        ),
                     }
                 )
     return spans
 
 
+def _status_message_and_events(
+    status: Optional[Dict[str, Any]], events: Any, decode_attrs: Any, time_key: str
+) -> Dict[str, Any]:
+    """A span's status message and events, when it has them (exceptions live in events).
+
+    OpenTelemetry records an exception as an ``exception`` event whose attributes
+    carry ``exception.type`` and ``exception.message``; the status message holds
+    the error description. Both are what a reader needs to see *why* a span failed.
+    """
+    out: Dict[str, Any] = {}
+    message = (status or {}).get("message") if isinstance(status, dict) else None
+    if isinstance(message, str) and message:
+        out["status_message"] = message
+    decoded = [
+        {
+            "name": event.get("name") or "event",
+            "time": _proto_ts_to_unix(event.get(time_key)),
+            "attributes": decode_attrs(event.get("attributes")),
+        }
+        for event in events or []
+        if isinstance(event, dict)
+    ]
+    if decoded:
+        out["events"] = decoded
+    return out
+
+
 def _decode_attrs(attrs: Optional[List[Dict[str, Any]]]) -> Dict[str, Any]:
     out: Dict[str, Any] = {}
     for a in attrs or []:
@@ -672,6 +702,9 @@ def _spans_from_otlp_dict(data: Dict[str, Any]) -> List[Dict[str, Any]]:
                         "end_time": _proto_ts_to_unix(span.get("end_time_unix_nano")),
                         "status": _status_from_dict(span.get("status")),
                         "attributes": attrs,
+                        **_status_message_and_events(
+                            span.get("status"), span.get("events"), _attrs_from_list, "time_unix_nano"
+                        ),
                     }
                 )
     return spans
diff --git a/simpleaudit/tracing/selection.py b/simpleaudit/tracing/selection.py
index c6332f0..c8c0ba8 100644
--- a/simpleaudit/tracing/selection.py
+++ b/simpleaudit/tracing/selection.py
@@ -569,8 +569,61 @@ def evidence_spans_for_turn(
     return result.selected
 
 
-def summarize_for_judge(result: SelectionResult, *, max_chars_per_span: int = 2000) -> str:
-    """Render selected spans into a compact text block for the judge prompt."""
+def _clip_values(value: Any, limit: int) -> Any:
+    """``value`` with every string longer than ``limit`` shortened (structure kept)."""
+    if isinstance(value, str):
+        return value if len(value) <= limit else value[:limit] + "…(truncated)"
+    if isinstance(value, dict):
+        return {key: _clip_values(item, limit) for key, item in value.items()}
+    if isinstance(value, list):
+        return [_clip_values(item, limit) for item in value]
+    return value
+
+
+def _span_body(attrs: Dict[str, Any], max_chars_per_span: int, max_chars_per_value: int) -> str:
+    """The span's attributes as JSON within ``max_chars_per_span``, every key kept.
+
+    Long values are shortened first (halving the per-value limit until the span
+    fits, down to 80 chars), so a long input history no longer pushes later
+    attributes such as the model's output out of view. Only if the keys alone
+    exceed the budget is the JSON cut at the end.
+    """
+    limit = max(80, max_chars_per_value)
+    body = json.dumps(_clip_values(attrs, limit), ensure_ascii=False, default=str)
+    while len(body) > max_chars_per_span and limit > 80:
+        limit = max(80, limit // 2)
+        body = json.dumps(_clip_values(attrs, limit), ensure_ascii=False, default=str)
+    if len(body) > max_chars_per_span:
+        body = body[:max_chars_per_span] + "…(truncated)"
+    return body
+
+
+def _failure_lines(span: Dict[str, Any], limit: int) -> List[str]:
+    """Why a span failed: its status message and exception events."""
+    lines: List[str] = []
+    if str(span.get("status") or "").upper() == "ERROR":
+        message = span.get("status_message")
+        lines.append("status: ERROR" + (f" — {str(message)[:limit]}" if message else ""))
+    for event in span.get("events") or []:
+        if not isinstance(event, dict) or event.get("name") != "exception":
+            continue
+        attrs = event.get("attributes") or {}
+        kind = attrs.get("exception.type") or "exception"
+        message = str(attrs.get("exception.message") or "")[:limit]
+        lines.append(f"exception: {kind}" + (f": {message}" if message else ""))
+    return lines
+
+
+def summarize_for_judge(
+    result: SelectionResult, *, max_chars_per_span: int = 2000, max_chars_per_value: int = 600
+) -> str:
+    """Render selected spans into a compact text block for the judge prompt.
+
+    Each span keeps all its attributes within ``max_chars_per_span``: long values
+    are shortened (``max_chars_per_value``, reduced further when needed) instead of
+    cutting the span's JSON from the front. A failed span also shows its status
+    message and exception.
+    """
     if not result.selected:
         note = f"\n(elided {result.elided_count} span(s))" if result.elided else ""
         return f"(no evidence spans){note}"
@@ -579,10 +632,9 @@ def summarize_for_judge(result: SelectionResult, *, max_chars_per_span: int = 20
         prov = span.get("provenance", {})
         header = f"[{prov.get('kind', '?')} {prov.get('name', '?')}] trace={prov.get('trace_id', '?')[:8]} span={prov.get('span_id', '?')[:8]}"
         attrs = _redacted(span.get("attributes") or {})
-        body = json.dumps(attrs, ensure_ascii=False, default=str)
-        if len(body) > max_chars_per_span:
-            body = body[:max_chars_per_span] + "…(truncated)"
-        blocks.append(f"{header}\n{body}")
+        body = _span_body(attrs, max_chars_per_span, max_chars_per_value)
+        lines = [header, *(_redacted(line) for line in _failure_lines(span, max_chars_per_value)), body]
+        blocks.append("\n".join(lines))
     text = "\n\n".join(blocks)
     if result.elided:
         text += f"\n\n(elided {result.elided_count} additional span(s): " + ", ".join(
diff --git a/simpleaudit/tracing/store.py b/simpleaudit/tracing/store.py
index cbfe48a..bc85514 100644
--- a/simpleaudit/tracing/store.py
+++ b/simpleaudit/tracing/store.py
@@ -69,6 +69,8 @@ def _attr(*keys: str) -> Any:
         "start_time": raw.get("start_time"),
         "end_time": raw.get("end_time"),
         "status": raw.get("status") or "OK",
+        # Why a span failed: the status message and the exception event (if any).
+        **{key: raw[key] for key in ("status_message", "events") if raw.get(key)},
     }
 
 
diff --git a/tests/test_tracing.py b/tests/test_tracing.py
index e47f352..274a2fa 100644
--- a/tests/test_tracing.py
+++ b/tests/test_tracing.py
@@ -832,3 +832,52 @@ async def _run():
             shared.stop()
 
     asyncio.run(_run())
+
+
+from simpleaudit.tracing.store import normalize_span  # noqa: E402
+
+
+def test_parse_otlp_keeps_status_message_and_exception_events():
+    """Why a span failed lives in its status message and its exception event."""
+    payload = {"resourceSpans": [{"scopeSpans": [{"spans": [{
+        "traceId": "t" * 32, "spanId": "s" * 16, "name": "execute_tool acme_get_order",
+        "startTimeUnixNano": "1000000000", "endTimeUnixNano": "2000000000",
+        "status": {"code": 2, "message": "ValueError: No order found"},
+        "events": [{"name": "exception", "timeUnixNano": "1500000000", "attributes": [
+            {"key": "exception.type", "value": {"stringValue": "ValueError"}},
+            {"key": "exception.message", "value": {"stringValue": "No order found for 'ACME-9999'."}},
+        ]}],
+    }]}]}]}
+    (span,) = parse_otlp_json(payload)
+    assert span["status"] == "ERROR"
+    assert span["status_message"] == "ValueError: No order found"
+    assert span["events"] == [{"name": "exception", "time": 1.5, "attributes": {
+        "exception.type": "ValueError", "exception.message": "No order found for 'ACME-9999'."}}]
+    normalized = normalize_span(span)
+    assert normalized["status_message"] == span["status_message"] and normalized["events"] == span["events"]
+
+
+def test_parse_otlp_without_events_adds_no_keys():
+    (span,) = parse_otlp_json(_otlp_payload())[:1]
+    assert "events" not in span and "status_message" not in span
+
+
+def test_summarize_for_judge_shortens_values_and_keeps_every_attribute():
+    """A long input history no longer pushes the model's output out of the span."""
+    span = {"span_id": "a" * 16, "trace_id": "b" * 32, "name": "chat", "kind": "LLM", "status": "OK",
+            "attributes": {"gen_ai.input.messages": "x" * 5000, "gen_ai.output.messages": "the answer"}}
+    text = summarize_for_judge(select_spans([span]))
+    assert '"gen_ai.output.messages": "the answer"' in text
+    assert "…(truncated)" in text
+    assert len(text) < 2300
+
+
+def test_summarize_for_judge_shows_why_a_span_failed():
+    span = normalize_span({"span_id": "a" * 16, "trace_id": "b" * 32, "name": "execute_tool", "kind": "TOOL",
+                           "status": "ERROR", "status_message": "lookup failed",
+                           "events": [{"name": "exception", "attributes": {
+                               "exception.type": "ValueError", "exception.message": "No order found"}}],
+                           "attributes": {"gen_ai.tool.name": "acme_get_order"}})
+    text = summarize_for_judge(select_spans([span]))
+    assert "status: ERROR — lookup failed" in text
+    assert "exception: ValueError: No order found" in text