publish the accuracy that was measured

The recommended detection profile showed operators precision 0.614, recall
0.606 and F1 0.607. Those three numbers appear nowhere in this repository
except the file that publishes them and a test that pinned them as literal
strings. The only recorded evaluation of that model at that operating point —
tile 512, overlap 64, threshold 0.15, the exact key its promotion report
recommended — reported 0.590, 0.577 and 0.582. The published figures were about
two and a half points more flattering than anything that was measured, on the
profile labelled "aanbevolen", and the test made sure nobody would correct
them. They now carry the measured values.

Worse in kind: the conservative profile reported "gemeten achtergrondfouten 0".
Its nine-sample hard-negative matrix at threshold 0.35 recorded 198 background
detections with 55 in the worst sample. The one number that tells an operator
whether a high-precision model invents buildings on empty terrain said zero
where the evidence said 55.

Those zeros are not simply wrong everywhere, which is why the fix is not just a
number. The other two profiles genuinely produced zero — against a strict
pure-empty gate of three samples, a different and much weaker test than the
nine-sample hard-negative matrix. Printing 0, 0 and 55 side by side invites a
comparison the evidence does not support, so each profile now states its gate,
its background sample count and the evaluation behind its figures, and the
panel shows them.

A test refuses any published figure that does not appear in the evidence
record, with a negative control so it cannot pass by matching nothing. Pinning
the numbers as literal strings is what let an unsourced precision survive; that
assertion is gone.

Also ignoring .codex-artifacts/ — ~300 MB of the rejected SAM2 and
edge-alignment trials plus a deploy bundle. Kept on disk, out of the repository.
No credentials in it; the two token scripts generate from settings.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitignore

@@ -73,3 +73,9 @@ build/
 .DS_Store
 .vscode/
 .idea/
+
+# Local investigation scratch: the SAM2 and edge-alignment trials the README
+# records as rejected, plus their imagery and a deploy bundle (~300 MB). Kept
+# on disk because the rejections are worth being able to re-check, out of the
+# repository because none of it is an input to the product.
+/.codex-artifacts/

backend/tests/test_detection_profile_metrics_are_traceable.py

@@ -0,0 +1,78 @@
+"""Every accuracy figure shown to an operator must exist in the evidence record.
+
+The recommended profile published precision 0.6140895327792112, recall
+0.6062221049337548 and F1 0.6068607646002744. Those three numbers appear
+nowhere in this repository except the file that publishes them and the test
+that pinned them as literal strings. The only recorded evaluation of that
+model at that operating point — tile 512, overlap 64, threshold 0.15 — reported
+0.5898197518, 0.5769921004 and 0.5824578632, so the published figures were
+about two and a half points more flattering than anything that was measured,
+and a test guaranteed nobody would correct them.
+
+An operator cannot check a number that has no source. This test refuses one.
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[2]
+PROFILES = ROOT / "frontend" / "src" / "components" / "detection" / "detectionProfiles.ts"
+EVIDENCE = ROOT / "docs" / "CODEX_EXECUTION_LOG.md"
+
+METRIC_FIELDS = ("precision", "recall", "f1")
+# The log rounds; the source file may carry more digits of the same value.
+TOLERANCE = 1e-9
+
+
+def _published_metrics() -> list[tuple[str, str, float]]:
+    source = PROFILES.read_text(encoding="utf-8")
+    profiles = re.findall(r"id: '([^']+)',(.*?)\n  \},", source, re.S)
+    assert profiles, "no operator profiles found; the file shape changed"
+
+    published: list[tuple[str, str, float]] = []
+    for profile_id, body in profiles:
+        for field in METRIC_FIELDS:
+            match = re.search(rf"^\s*{field}: ([0-9.]+),", body, re.M)
+            assert match, f"{profile_id} publishes no {field}"
+            published.append((profile_id, field, float(match.group(1))))
+    return published
+
+
+def _recorded_values() -> list[float]:
+    text = EVIDENCE.read_text(encoding="utf-8")
+    return [float(value) for value in re.findall(r"\b0\.\d{4,}\b", text)]
+
+
+def test_every_published_accuracy_figure_appears_in_the_evidence_record() -> None:
+    recorded = _recorded_values()
+    untraceable = [
+        f"{profile_id}.{field} = {value}"
+        for profile_id, field, value in _published_metrics()
+        if not any(abs(value - candidate) <= TOLERANCE for candidate in recorded)
+    ]
+
+    assert not untraceable, (
+        "These figures are shown to operators but were never recorded in "
+        f"docs/CODEX_EXECUTION_LOG.md: {untraceable}. Publish the measurement "
+        "that was taken, or record the evaluation that produced these."
+    )
+
+
+def test_each_profile_names_the_measurement_behind_its_numbers() -> None:
+    source = PROFILES.read_text(encoding="utf-8")
+    profile_count = source.count("modelAssetId:")
+
+    assert source.count("evidenceReference:") == profile_count
+    assert source.count("backgroundGate:") == profile_count
+    assert source.count("backgroundSampleCount:") == profile_count
+
+
+def test_the_check_would_notice_an_invented_figure() -> None:
+    """Without this the test could pass because nothing ever matches."""
+
+    recorded = _recorded_values()
+
+    assert any(abs(0.5898197518 - value) <= TOLERANCE for value in recorded)
+    assert not any(abs(0.6140895327792112 - value) <= TOLERANCE for value in recorded)

backend/tests/test_sprint155_detection_operator_profiles.py

@@ -22,11 +22,12 @@ def test_detection_operator_profiles_define_explicit_historical_yolo_controls_wi
     assert "promotionRecommendation" not in source
     assert "independentTestProven: false" in source
     assert "positiveSampleCount: 7" in source
-    assert "precision: 0.6140895327792112" in source
-    assert "recall: 0.6062221049337548" in source
-    assert "f1: 0.6068607646002744" in source
-    assert "f1: 0.5432865390636915" in source
+    # The figures themselves are checked against the evidence record in
+    # tests/test_detection_profile_metrics_are_traceable.py. Pinning them as
+    # literals here is what kept an unsourced precision in the product.
     assert "maxBackgroundDetections: 0" in source
+    assert "maxBackgroundDetections: 55" in source
+    assert "backgroundGate:" in source
     assert "Slechts drie pure-achtergrondbeelden" in source
     assert "ruimtelijke onafhankelijkheid niet bewezen" in source
     assert "controlekandidaat en niet als grondwaarheid" in source

docs/KNOWN_LIMITATIONS.md

@@ -82,6 +82,14 @@ runtime source of truth.
   different pages is kept once instead, since a cursor over a changing table
   can produce that legitimately.
 
+- The operator profiles in Detection Lab publish figures from historical
+  evaluations, and each names the measurement it came from. Their background
+  numbers are not comparable across profiles: the two showing zero were checked
+  against a strict pure-empty gate of three samples, while the conservative
+  profile's 55 comes from a nine-sample hard-negative matrix that also recorded
+  198 background detections in total. None of the three has an independent
+  hold-out.
+
 ## Historical analysis
 
 - Historical comparison requires compatible source editions, units, coverage

frontend/src/components/detection/DetectionModelManagement.tsx

@@ -165,8 +165,14 @@ export function DetectionModelManagement({
                       <span>historische herkenningsgraad {profile.recall.toFixed(3)}</span>
                       <span>historische F1 {profile.f1.toFixed(3)}</span>
                       <span>positieve controles {profile.positiveSampleCount}</span>
-                      <span>gemeten achtergrondfouten {profile.maxBackgroundDetections}</span>
+                      <span>
+                        ergste achtergrondmonster {profile.maxBackgroundDetections} van{' '}
+                        {profile.backgroundSampleCount}
+                      </span>
                     </div>
+                    <p className="field-guidance">
+                      Achtergrondtoets: {profile.backgroundGate}. Meting: {profile.evidenceReference}.
+                    </p>
                     <p className="field-guidance">{profile.limitationMessage}</p>
                     <button
                       className="secondary-action"

frontend/src/components/detection/detectionProfiles.ts

@@ -9,11 +9,18 @@ export interface DetectionOperatorProfile {
   f1: number
   positiveSampleCount: number
   maxBackgroundDetections: number
+  backgroundSampleCount: number
+  backgroundGate: string
+  evidenceReference: string
   validationScope: string
   description: string
   limitationMessage: string
 }
 
+// Every number below is a value recorded in docs/CODEX_EXECUTION_LOG.md for
+// this exact model, tile size, overlap and threshold. Publishing a figure the
+// evidence record does not contain is the one thing an operator cannot check,
+// so tests/test_detection_profile_metrics_are_traceable.py refuses it.
 export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
   {
     id: 'small-building-balanced-review',
@@ -21,16 +28,19 @@ export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
     modelAssetId: 'geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt',
     confidenceThreshold: 0.15,
     independentTestProven: false,
-    precision: 0.6140895327792112,
-    recall: 0.6062221049337548,
-    f1: 0.6068607646002744,
+    precision: 0.5898197518,
+    recall: 0.5769921004,
+    f1: 0.5824578632,
     positiveSampleCount: 7,
     maxBackgroundDetections: 0,
+    backgroundSampleCount: 3,
+    backgroundGate: 'strikte lege-poort (Postel, Lommel, Arendonk)',
+    evidenceReference: 'promotierapport smallbld-minpx3-img640-ft30 | tegel 512 | overlap 64 | drempel 0.15',
     validationScope: 'historische operatorcontrole in Mol en de Kempen; ruimtelijke onafhankelijkheid niet bewezen',
     description:
       'Historisch controleprofiel met een evenwicht tussen gevonden en gemiste kleine gebouwen. Gebruik dit als startpunt voor lokale QA, niet als vrijgavebewijs.',
     limitationMessage:
-      'Slechts drie pure-achtergrondbeelden en geen onafhankelijke hold-out ondersteunen deze historische meting. Postel bleef het moeilijkste gebied; behandel elke detectie als controlekandidaat en niet als grondwaarheid.',
+      'Slechts drie pure-achtergrondbeelden en geen onafhankelijke hold-out ondersteunen deze historische meting. Dit profiel ruilt precisie in voor dekking: het vindt meer kleine gebouwen dan het vorige, maar levert een hogere controlelast op. Postel bleef het moeilijkste gebied; behandel elke detectie als controlekandidaat en niet als grondwaarheid.',
   },
   {
     id: 'expanded-balanced-review',
@@ -38,11 +48,14 @@ export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
     modelAssetId: 'geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt',
     confidenceThreshold: 0.15,
     independentTestProven: false,
-    precision: 0.6470590036169351,
-    recall: 0.4699913836847832,
-    f1: 0.5432865390636915,
+    precision: 0.6470590036,
+    recall: 0.4699913837,
+    f1: 0.5432865391,
     positiveSampleCount: 7,
     maxBackgroundDetections: 0,
+    backgroundSampleCount: 3,
+    backgroundGate: 'strikte lege-poort (Postel, Lommel, Arendonk)',
+    evidenceReference: 'zeven-AOI positieve matrix aoi1024expandedminpx4vis035e50 | tegel 512 | overlap 64 | drempel 0.15',
     validationScope: 'historische operatorcontrole in Mol en de Kempen; ruimtelijke onafhankelijkheid niet bewezen',
     description: 'Historisch profiel voor lokale controles waarbij minder foutieve vondsten belangrijker zijn dan maximale dekking.',
     limitationMessage:
@@ -58,10 +71,13 @@ export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
     recall: 0.202135,
     f1: 0.32086574003576274,
     positiveSampleCount: 7,
-    maxBackgroundDetections: 0,
+    maxBackgroundDetections: 55,
+    backgroundSampleCount: 9,
+    backgroundGate: 'hard-negative matrix (negen monsters, 198 detecties in totaal)',
+    evidenceReference: 'conservatieve zeven-AOI matrix + negen-monster hard-negative matrix aoi1024bg512r3e50 | drempel 0.35',
     validationScope: 'historische operatorcontrole in Mol en de Kempen; ruimtelijke onafhankelijkheid niet bewezen',
     description: 'Historisch profiel voor lokale controles waarbij zo weinig mogelijk foutieve vondsten zwaarder wegen dan volledige dekking.',
     limitationMessage:
-      'De historische precisie gaat samen met zeer lage herkenningsgraad. Resultaten blijven controlebewijs en geen automatische waarheid of releasebewijs.',
+      'De historische precisie gaat samen met zeer lage herkenningsgraad. De hoge precisie geldt binnen bebouwde AOI-gebieden: op de negen hard-negative monsters produceerde dit profiel nog 198 achtergronddetecties, met 55 in het zwaarste monster. Resultaten blijven controlebewijs en geen automatische waarheid of releasebewijs.',
   },
 ]