diff --git a/graphify/build.py b/graphify/build.py
index caeaefdf9..bc8fce1d6 100644
--- a/graphify/build.py
+++ b/graphify/build.py
@@ -30,7 +30,7 @@
 import networkx as nx
 from .ids import make_id, normalize_id as _normalize_id
 from .paths import default_graph_json as _default_graph_json
-from .validate import validate_extraction
+from .validate import validate_extraction, VALID_FILE_TYPES
 
 
 # Language interop families, keyed by extension, for the cross-language phantom-edge
@@ -72,6 +72,20 @@
     "framework": "concept",
 }
 
+# Biomedical entity categories that semantic subagents are told to record in a
+# node's `entry_type` (see the biomedical extraction rules in the /graphify
+# skill). Models sometimes leak the category into `file_type` instead, where it
+# is not a valid file_type. Unlike the generic synonyms above — which collapse
+# unknown types to `concept` and discard the original token — these categories
+# carry real information, so we preserve them in `entry_type` rather than lose
+# them. `technology` is intentionally also present in _FILE_TYPE_SYNONYMS (it is
+# ambiguous with the code-corpus sense), so it is only recovered as biomedical
+# when the node already shows a biomedical signal (an existing `entry_type`).
+_BIOMEDICAL_CATEGORIES = {
+    "drug", "gene", "disease", "protein", "technology",
+    "peptide", "institution", "company",
+}
+
 
 # Hyperedge member lists are canonically keyed `nodes` (see graphify/llm.py
 # extraction spec), but LLM/subagent drift and externally-supplied graph.json
@@ -428,8 +442,19 @@ def build_from_json(extraction: dict, *, directed: bool = False, root: str | Pat
         if node.get("file_type") in (None, ""):
             node["file_type"] = "concept"
         ft = node.get("file_type", "")
-        if ft and ft not in {"code", "document", "paper", "image", "rationale", "concept"}:
-            node["file_type"] = _FILE_TYPE_SYNONYMS.get(ft, "concept")
+        if ft and ft not in VALID_FILE_TYPES:
+            # A leaked biomedical entity category (drug/gene/disease/...) carries
+            # real information, so preserve it in entry_type instead of
+            # collapsing it to `concept` and discarding it. `technology` is
+            # ambiguous with the code-corpus synonym, so only treat it as
+            # biomedical when the node already carries a biomedical signal.
+            if ft in _BIOMEDICAL_CATEGORIES and (
+                node.get("entry_type") or ft not in _FILE_TYPE_SYNONYMS
+            ):
+                node.setdefault("entry_type", ft)
+                node["file_type"] = "document"
+            else:
+                node["file_type"] = _FILE_TYPE_SYNONYMS.get(ft, "concept")
 
     # Canonicalize hyperedge member lists (#1561): producers sometimes key the
     # member list `members`/`node_ids` instead of `nodes`. Fold aliases onto
diff --git a/tests/test_build.py b/tests/test_build.py
index ed87e6b11..1c9bbf9b6 100644
--- a/tests/test_build.py
+++ b/tests/test_build.py
@@ -994,3 +994,55 @@ def test_build_from_json_prunes_dangling_hyperedge_members(capsys):
     assert set(hes) == {"he_partial"}, "an all-dangling hyperedge must be dropped"
     assert hes["he_partial"]["nodes"] == ["alpha", "beta"]
     assert "he_all_ghost" in capsys.readouterr().err
+
+
+def test_leaked_biomedical_category_recovered_into_entry_type():
+    # #(new): biomedical semantic subagents are told to put the entity category
+    # (drug/gene/disease/...) in `entry_type`, but models sometimes also write
+    # it to `file_type`, which is not a valid file_type. Recover the category
+    # into entry_type and coerce file_type to a valid `document` — without
+    # emitting an invalid-file_type warning and without discarding the category.
+    ext = {
+        "nodes": [
+            # peptide is biomedical-exclusive: recovered even without entry_type
+            {"id": "phlip", "label": "pHLIP", "file_type": "peptide",
+             "source_file": "patents/records/x.md"},
+            # entry_type already correct; must not be overwritten
+            {"id": "mitocatch", "label": "MitoCatch", "file_type": "technology",
+             "entry_type": "technology", "source_file": "patents/records/y.md"},
+        ],
+        "edges": [],
+    }
+    G = build_from_json(ext)
+    assert G.nodes["phlip"]["file_type"] == "document"
+    assert G.nodes["phlip"]["entry_type"] == "peptide"
+    assert G.nodes["mitocatch"]["file_type"] == "document"
+    assert G.nodes["mitocatch"]["entry_type"] == "technology"
+
+
+def test_ambiguous_technology_without_signal_stays_code_synonym():
+    # `technology` is ambiguous with the code-corpus synonym (-> concept). With
+    # no biomedical signal (no entry_type), preserve the existing behavior so
+    # non-biomedical corpora are unaffected.
+    ext = {
+        "nodes": [
+            {"id": "n1", "label": "Some Tech", "file_type": "technology",
+             "source_file": "docs/a.md"},
+        ],
+        "edges": [],
+    }
+    G = build_from_json(ext)
+    assert G.nodes["n1"]["file_type"] == "concept"
+    assert "entry_type" not in G.nodes["n1"]
+
+
+def test_unknown_invalid_file_type_still_falls_back_to_concept():
+    ext = {
+        "nodes": [
+            {"id": "n1", "label": "Junk", "file_type": "gizmo",
+             "source_file": "docs/a.md"},
+        ],
+        "edges": [],
+    }
+    G = build_from_json(ext)
+    assert G.nodes["n1"]["file_type"] == "concept"