diff --git a/py/api/admin.py b/py/api/admin.py
index 1bf0124a..ec3f6db7 100644
--- a/py/api/admin.py
+++ b/py/api/admin.py
@@ -1758,14 +1758,18 @@ class AdminRoutesMixin:
             # No node takes a positive or negative input (a guider or wrapper
             # graph): the one wired text, never a guess (#252).
             found = sampler_less_prompt(canonical["prompt"])
-            if found["state"] != "resolved":
+            prompt_text = found["text"]
+            if prompt_text is None and isinstance(canonical.get("parameters"), str):
+                # A metadata saver also wrote A1111 text: its positive prompt.
+                parsed = self._parse_comfyui_prompt(metadata)["positive_prompt"]
+                prompt_text = parsed.strip() if parsed and parsed.strip() else None
+            if prompt_text is None:
                 linked = self._link_recorded_run(
                     media_file, canonical, detailed, history
                 )
                 if linked is not None:
                     return linked
                 return "unresolved" if found["state"] == "ambiguous" else "found"
-            prompt_text = found["text"]
         else:
             prompt_text = self._parse_comfyui_prompt(metadata)["positive_prompt"]
             if not (prompt_text and prompt_text.strip()):
diff --git a/tests/test_pair_scan.py b/tests/test_pair_scan.py
index c896e588..2ee70151 100644
--- a/tests/test_pair_scan.py
+++ b/tests/test_pair_scan.py
@@ -410,5 +410,42 @@ class TestSamplerLessScan(AdminAPITestCase):
                 {"id": 6, "type": "CLIPTextEncode", "widgets_values": ["stale text"]}
             ]
         }
-        self.assertIn(self._scanned("stale.png", g, workflow), ("found", "unresolved"))
+        self.assertEqual(self._scanned("stale.png", g, workflow), "unresolved")
         self.assertEqual(self._texts(), [])
+
+    def test_a_lora_loaders_trigger_word_output_is_not_its_text(self):
+        # The encoder reads the loader's trigger-word output (slot 2), not its
+        # text input: the prompt is unknown, never the "<lora:...>" syntax.
+        g = flux_graph(text=["50", 2])
+        g["50"] = {
+            "class_type": "Lora Loader (LoraManager)",
+            "inputs": {"model": ["1", 0], "text": "<lora:style:1.00>"},
+        }
+        g["22"]["inputs"]["model"] = ["50", 0]
+        self.assertEqual(self._scanned("trigger.png", g), "unresolved")
+        self.assertEqual(self._texts(), [])
+
+    def test_a_seeded_text_node_holds_a_template_not_the_prompt(self):
+        g = flux_graph(text=["80", 0])
+        g["80"] = {
+            "class_type": "RandomPrompt",
+            "inputs": {"text": "a {red|blue} fox", "seed": 5},
+        }
+        self.assertEqual(self._scanned("random.png", g), "unresolved")
+        self.assertEqual(self._texts(), [])
+
+    def test_a1111_parameters_beside_an_undecided_graph_are_read(self):
+        # Metadata savers write both; the old scan read the A1111 positive.
+        g = flux_graph()
+        g["50"] = {
+            "class_type": "Lora Loader (LoraManager)",
+            "inputs": {"model": ["1", 0], "text": "<lora:style:1.00>"},
+        }
+        g["22"]["inputs"]["model"] = ["50", 0]
+        meta = {
+            "prompt": json.dumps(g),
+            "parameters": "a red fox\nNegative prompt: blur\nSteps: 20",
+        }
+        path = self.output_dir / "saver.png"
+        self.assertEqual(self.api._ingest_scanned_file(path, meta), "added")
+        self.assertEqual(self._texts(), ["a red fox"])
diff --git a/tests/test_sampler_less_prompt.py b/tests/test_sampler_less_prompt.py
index ca8d8b5d..7f6379ab 100644
--- a/tests/test_sampler_less_prompt.py
+++ b/tests/test_sampler_less_prompt.py
@@ -31,6 +31,22 @@ class TestSamplerLessPrompt(unittest.TestCase):
             graph = _guided(**{"1": {"class_type": "Enc", "inputs": {"text": value}}})
             self.assertEqual(sampler_less_prompt(graph)["state"], "absent", value)
 
+    def test_number_like_words_are_prompts(self):
+        for value in ("nan", "inf", "Infinity", "1_000"):
+            graph = _guided(**{"1": {"class_type": "Enc", "inputs": {"text": value}}})
+            self.assertEqual(sampler_less_prompt(graph)["text"], value)
+
+    def test_an_own_node_with_unreadable_text_beside_another_text_is_ambiguous(self):
+        graph = _guided(
+            **{
+                "1": {"class_type": "PromptManager", "inputs": {"text": ["5", 0]}},
+                "5": {"class_type": "LLM", "inputs": {"seed": 3}},
+                "2": {"class_type": "Enc", "inputs": {"text": "<lora:x:1>"}},
+                "3": {"class_type": "Sink", "inputs": {"c": ["2", 0]}},
+            }
+        )
+        self.assertEqual(sampler_less_prompt(graph)["state"], "ambiguous")
+
     def test_unhashable_class_type_is_an_unknown_node_not_a_crash(self):
         graph = _guided(**{"1": {"class_type": ["x"], "inputs": {"text": "a fox"}}})
         self.assertEqual(sampler_less_prompt(graph)["text"], "a fox")
diff --git a/utils/prompt_pair_resolver.py b/utils/prompt_pair_resolver.py
index b36ac0a2..5ff4324a 100644
--- a/utils/prompt_pair_resolver.py
+++ b/utils/prompt_pair_resolver.py
@@ -2580,13 +2580,12 @@ def resolve_text_input(
         return None
 
 
+_NUMBER = re.compile(r"[+-]?(?:\d+\.?\d*|\.\d+)(?:[eE][+-]?\d+)?")
+
+
 def _numeric(text: str) -> bool:
-    """Whether ``text`` is a number (a widget value such as ``2048``, #252)."""
-    try:
-        float(text)
-    except ValueError:
-        return False
-    return True
+    """Whether ``text`` is a plain number (a widget value such as ``2048``)."""
+    return _NUMBER.fullmatch(text) is not None
 
 
 def sampler_less_prompt(graph: Any, *, max_visits: int = 500) -> Dict[str, Any]:
@@ -2599,7 +2598,8 @@ def sampler_less_prompt(graph: Any, *, max_visits: int = 500) -> Dict[str, Any]:
     feeds nothing did not make the image. Its text is traced as a STRING
     (``resolve_text_input``); a linked text no rule reads is unknown unless
     it is built from a candidate's text (``_covered``). Blank and
-    numeric texts are never prompts. This repository's own positive-text
+    numeric texts are never prompts, and a node with a seed input holds a
+    template, not the prompt (unknown). This repository's own positive-text
     nodes are authoritative: when they give a text, only they count.
 
     Returns ``{"state", "text", "texts"}``: ``resolved`` with the one distinct
@@ -2624,11 +2624,16 @@ def sampler_less_prompt(graph: Any, *, max_visits: int = 500) -> Dict[str, Any]:
         inputs = node.get("inputs") if isinstance(node, dict) else None
         if node_id not in consumed or not _has_text_input(node):
             continue
+        if _seeded(inputs):
+            # Generated at run time (a wildcard or a language model): its
+            # text input is a template, not the prompt that ran.
+            unknown = True
+            continue
         value = inputs["text"]
         text = resolve_text_input(nodes, node_id, "text", max_visits=max_visits)
         if text is None:
             # A literal that is not a string (a number) is never a prompt.
-            if _link(value) and not _covered(nodes, str(value[0])):
+            if _link(value) and not _covered(nodes, value):
                 unknown = True
             continue
         text = text.strip()
@@ -2652,26 +2657,36 @@ def _has_text_input(node: Any) -> bool:
     return isinstance(inputs, dict) and "text" in inputs
 
 
-def _covered(nodes: dict, source_id: str) -> bool:
+def _seeded(inputs: dict) -> bool:
+    """A seed input: the node's string output is a run-time value (the
+    resolver's ``runtime-value`` rule)."""
+    return any(isinstance(name, str) and name.endswith("seed") for name in inputs)
+
+
+def _covered(nodes: dict, link: list) -> bool:
     """Whether an unreadable text link is built from a candidate's text.
 
-    True when a node upstream of ``source_id`` (itself included) has a
-    ``text`` input: that candidate is read on its own, and whatever else is
-    joined to it at run time (a caption, a language model) is no stored
-    prompt to choose. False when the text comes only from inputs no candidate
-    holds (a builder's ``prompt`` field): the prompt is unknown then.
+    True when a node upstream of ``link`` is read through its first output
+    (slot 0, the legacy encoder rule's text) and has an unseeded ``text``
+    input: that candidate is read on its own, and whatever else is joined to
+    it at run time (a caption, a language model) is no stored prompt to
+    choose. Another output of such a node (a loader's trigger words) is not
+    its text. False when the text comes only from inputs no candidate holds
+    (a builder's ``prompt`` field): the prompt is unknown then.
     """
-    seen, stack = set(), [source_id]
+    seen, stack = set(), [(str(link[0]), link[1])]
     while stack:
-        node_id = stack.pop()
-        if node_id in seen:
+        node_id, slot = stack.pop()
+        if (node_id, slot) in seen:
             continue
-        seen.add(node_id)
+        seen.add((node_id, slot))
         node = nodes.get(node_id)
         inputs = node.get("inputs") if isinstance(node, dict) else None
         if not isinstance(inputs, dict):
             continue
-        if "text" in inputs:
+        if "text" in inputs and slot == 0 and not _seeded(inputs):
             return True
-        stack.extend(str(value[0]) for value in inputs.values() if _link(value))
+        stack.extend(
+            (str(value[0]), value[1]) for value in inputs.values() if _link(value)
+        )
     return False