diff --git a/observe/extract.py b/observe/extract.py
index ffd27a2d6..0c80cb9d4 100644
--- a/observe/extract.py
+++ b/observe/extract.py
@@ -22,6 +22,10 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+_SCHEMA = json.loads(
+    (Path(__file__).parent / "extract.schema.json").read_text(encoding="utf-8")
+)
+
 # Default maximum frames to extract content from
 DEFAULT_MAX_EXTRACTIONS = 20
 
@@ -244,6 +248,7 @@ def _ai_select_frames(
         context="observe.extract.selection",
         system_instruction=prompt_content.text,
         json_output=True,
+        json_schema=_SCHEMA,
         thinking_budget=4096,
         max_output_tokens=1024,
         temperature=0.3,
diff --git a/observe/extract.schema.json b/observe/extract.schema.json
new file mode 100644
index 000000000..e8fca4e56
--- /dev/null
+++ b/observe/extract.schema.json
@@ -0,0 +1,5 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "type": "array",
+  "items": {"type": "integer", "minimum": 0}
+}
diff --git a/tests/test_extract_schema.py b/tests/test_extract_schema.py
new file mode 100644
index 000000000..a68c0015c
--- /dev/null
+++ b/tests/test_extract_schema.py
@@ -0,0 +1,60 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright (c) 2026 sol pbc
+
+import importlib
+import json
+from pathlib import Path
+
+from jsonschema import Draft202012Validator
+
+import think.models as models
+
+extract_mod = importlib.import_module("observe.extract")
+
+_SCHEMA = json.loads(
+    (Path(__file__).resolve().parents[1] / "observe" / "extract.schema.json").read_text(
+        encoding="utf-8"
+    )
+)
+
+
+def test_extract_schema_file_is_valid_draft_2020_12():
+    Draft202012Validator.check_schema(_SCHEMA)
+
+
+def test_extract_schema_accepts_and_rejects_expected_values():
+    validator = Draft202012Validator(_SCHEMA)
+
+    assert validator.is_valid([])
+    assert validator.is_valid([1, 15, 42, 89])
+    assert validator.is_valid([1, 0])
+    assert not validator.is_valid(["1"])
+    assert not validator.is_valid([-1])
+    assert not validator.is_valid([1.5])
+    assert not validator.is_valid(42)
+    assert not validator.is_valid({"ids": [1]})
+    assert not validator.is_valid([[1, 2]])
+
+
+def test_ai_select_frames_passes_schema_to_generate(monkeypatch):
+    captured = {}
+
+    def fake_generate(**kwargs):
+        captured.update(kwargs)
+        return "[1]"
+
+    monkeypatch.setattr(models, "generate", fake_generate)
+
+    frames = [
+        {"frame_id": 1, "timestamp": 1.0, "analysis": {"primary": "code"}},
+    ]
+    categories = {"code": {"description": "Code editors"}}
+
+    result = extract_mod._ai_select_frames(
+        frames,
+        max_extractions=5,
+        categories=categories,
+    )
+
+    assert captured["json_schema"] is extract_mod._SCHEMA
+    assert result == [1]