JensandClaude Opus 5
5c509ca2a9
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>
79 lines
3.1 KiB
Python
79 lines
3.1 KiB
Python
"""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)
|