scripts/evaluate_belgium_building_candidate.py freezes its post-processing
before the protected test — NMS IoU and a containment threshold selected during
calibration, defaulting to 1.0. The runtime applied a hardcoded 0.85, so a
promoted candidate was served under suppression its evaluation never measured
and dropped detections the gate had counted. Neither report showed the
difference. That constant was mine, added without noticing the evaluation
pipeline already had a tuned value for the same concept.
Containment is now configuration, recorded on every run beside the duplicate
IoU threshold, so an operator can serve a candidate at the value it was gated
at and afterwards see which value produced a given score.
Two runs that post-processed differently produced different candidate sets from
the same model output, so the comparison endpoint refuses to rank them. Runs
recorded before those values were persisted carry none, and absence is not
treated as a difference.
The remaining gap is deliberate and documented rather than closed: the gate
scores the model on its protected test set, the workbench scores the whole
pipeline including coverage clipping and the tile-edge filter. A promoted
candidate will not reproduce its gate F1 exactly, and pretending otherwise
would be the worse answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Detection QA reports a precision/recall curve, average precision and a
calibration sweep; segmentation QA reported a single operating point. Both rank
their outputs by confidence, so the same view applies, and the asymmetry meant
the two panels answered different questions about comparable runs — an
inconsistency introduced when detection gained the curve.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The review vocabulary already separates a model error from a reference gap,
because the product's position is that official footprints are not
automatically perfect ground truth. Those verdicts were only counted. An
operator who inspected forty false positives and established that twelve are
buildings the reference simply lacks still saw a precision counting all forty
against the model — a number they had personally disproved, on the panel where
they disproved it.
Applying the verdicts gives an adjudicated score reported next to the raw one,
so nothing is quietly improved. Not being able to judge is not evidence in the
model's favour, so uncertain and obscured verdicts keep counting, as does a
decision from a later release that this runtime does not recognise.
Because part of the evidence is usually still unreviewed, the honest form is an
interval rather than a single corrected number: pessimistic assumes every
unreviewed finding is a model error, optimistic assumes none is, and the
headline equals the pessimistic reading so a partly reviewed check never
presents as a settled one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The workbench ranks model variants by a stored F1, each measured at that
variant's own confidence threshold. A conservatively calibrated detector then
looks worse than a liberal one without detecting anything differently: the
number says as much about the threshold as about the model. POST
/detection/runs/compare ranks on average precision instead, which describes the
whole ranking a model produced, and keeps each run's own-threshold F1 visible
next to it so the difference between the two readings is auditable.
Comparability comes before the ranking. Runs over different source rasters,
scored against different references, without a proven inference footprint, or
covering a different evaluated population are not alternatives to one another,
and no metric makes them so. The report names which of those applies and still
returns the numbers — they are simply not a ranking.
Each run is scored through the same QA path the workbench uses, so a comparison
and the persisted quality checks cannot drift apart.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Threshold calibration ran the model over every tile once per threshold — three
GPU passes to compare 0.50, 0.25 and 0.15 on a hundred-tile raster. The answer
is already in a single run at the lowest value: detections above a higher cut
are a subset of it, and duplicate suppression walks candidates in descending
confidence, so a lower-confidence box can never displace a higher-confidence
one. The kept set above any cut is identical whichever threshold the run used,
which is what makes one pass sufficient rather than merely cheaper.
QA now takes calibration_thresholds and reads each operating point off the same
precision/recall walk it already performs, marking the F1-optimal cut. The lab
runs inference once and fills its table from the sweep.
The contract test asserted the per-threshold loop by name, pinning the waste it
was meant to describe. It now states what calibration owes an operator: a row
per requested threshold, from one run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
tile_manifest_path arrives in the detection and segmentation request and was
read straight off disk, and a manifest entry may name an absolute tile path.
That makes an API field an unbounded reference to the host filesystem, and it
contradicts the rule the persistence model rests on: only a governed,
runtime-produced artifact may be consumed, and a file outside the storage root
is not one.
Both the manifest and every tile it names now resolve under STORAGE_ROOT.
Resolution happens before the comparison, so ".." cannot climb out and a
sibling that merely shares a name prefix does not pass.
GEOINTEL_ALLOW_EXTERNAL_ARTIFACT_PATHS opts out for provisioning workflows that
stage tiles before ingest.
The check honours the Settings the caller is operating under rather than the
process-wide ones, because every analysis path already threads its own.
The affected tests write manifests into tmp_path, so they now declare tmp_path
as the storage root — which is what a deployment does, and makes the fixtures
more honest than they were.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
/detection/runs/{id}/detections and its GeoJSON sibling returned every
persisted detection, as did the segmentation equivalents. A regional run holds
tens of thousands, and these are the endpoints the results table and the map
overlay call after every run.
They now take limit and offset, default to 2.000, and report total, limit,
offset and truncated so the complete population stays visible while what is
transferred does not. The GeoJSON responses carry the same window in a
geointel_result_window foreign member.
Rows are ordered by confidence, so a capped overlay draws the strongest
detections rather than an arbitrary slice, and the lab says how many of how
many are being shown rather than silently presenting a page as the whole run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
evidence_geojson emitted one feature per false positive, one per false negative
and two per match, with no limit. A regional check of 40k detections against 45k
reference footprints produced well over a hundred thousand features in a single
response, plus one warning string per unresolvable identifier. The endpoint the
entire review workflow depends on therefore failed exactly where review matters
most.
What to draw is now decided before any geometry is fetched, so the query work is
proportional to the result rather than to the size of the check — previously
130k geometries were resolved through an IN clause holding every identifier in
the check, to then discard most of them.
The budget is split between misses and false positives in proportion to their
populations with at least one of each, rather than by strict priority, which
would mean a check with 50.000 misses and three false positives never showed
one. Confirmations fill what remains, and a match is kept or dropped as a pair
because half a match is not reviewable evidence.
limit_evidence and evidence_role_counts are removed: plan_evidence supersedes
them, and helpers kept alive only by their own tests read like a contract.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The README's architecture diagram and stack list promised Redis + RQ. There is
no such thing in this codebase: redis_url is a settings field nothing reads,
rq is not a dependency, and the only worker is an asyncio poller over a
Postgres job table. Describe what exists.
API_CONTRACTS and KNOWN_LIMITATIONS gain the rules a reader needs to interpret
a result: deterministic confidence-ranked matching, the precision/recall curve,
the three flood-hazard cell populations, the sub-cell selection fallback, the
whole-feature object count next to clipped area metrics, and the required
manifest CRS. The new limitations are stated as limits rather than buried:
raster analysis has no sub-cell precision, and the object count is not
fractional.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Jens