Detection and segmentation run listings returned every run a project had ever
produced. Runs accumulate with every analysis while the panel only ever draws
the recent ones, so the response grew without bound for no benefit.
Both take limit and offset now and report total, limit, offset and truncated,
matching the result listings.
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>
/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>
Change detection was the one analysis that ignored the selection entirely. It
compared two datasets in full, loaded every feature of both into Python with no
spatial predicate, and — with include_unchanged defaulting to true — returned a
FeatureCollection holding both datasets. For a regional building layer that is
the wrong answer to "what changed here" and a response no browser should be
asked to hold.
It now accepts bbox and area_id, resolved the way every other analysis resolves
them, and loads through an indexed ST_Intersects predicate.
Features are deliberately not clipped to the selection. A change class
describes a whole object: comparing a clipped earlier footprint against an
unclipped later one would report the selection edge itself as a change. Objects
the edge crosses are compared in full and counted in a warning.
The returned geometry is capped by preview_limit, spending that budget on
modified, added and removed before unchanged, while every count still describes
the whole selection.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
disclosure from data
Change detection had only added/removed/unchanged, so a building extended by
an annexe dropped below the IoU threshold and was reported twice: once as
removed and once as added. That hides exactly the category a change-detection
product exists to show and inflates both counts. A "modified" class now covers
the band between the modified floor and the unchanged threshold.
Matching also ran as a full cross product with no spatial index, unlike the QA
matcher beside it: two municipal building layers meant hundreds of millions of
geometry intersections. It uses an STRtree and considers larger footprints
first, so a big footprint is not left over after a small neighbour claimed its
counterpart.
The assistant guaranteed honesty about estimated values by rewriting the
model's sentences with regular expressions, which only fires when it
recognises the phrasing the model happened to produce. estimate_disclosures
derives the same statement from the metric metadata, so it holds regardless of
how the answer was worded. The prose substitution stays as a second layer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Four ways a selection produced a confident number about a different area than
the operator drew:
Flood hazard divided the inundated cells by every cell in the drawn rectangle,
including cells the VMM raster does not model at all. A selection reaching
past the modelled extent therefore reported a diluted risk share, turning
missing data into an implied absence of risk. Terrain, bathymetry and thematic
raster already divided by valid cells; flood hazard was the outlier. It now
reports the three populations separately, states model coverage next to the
drawn area, and returns a null fraction rather than a zero when nothing was
modelled.
geometry_mask selects a cell when its centre falls inside the geometry, so a
rectangle smaller than one cell — or one landing between four centres —
selected nothing and the analysis returned zeros indistinguishable on screen
from "we looked and there is nothing here". On a 100 m population raster a
40 m rectangle over a city block reported no inhabitants. Selection now falls
back to the touched cells and says that it did, since the answer then covers
more ground than was requested. rasterio.mask applies the same centre rule
when cropping, so that call is widened too; the cells that count are still
decided by the centre rule wherever it selects anything.
The object count treated any feature touching the selection as whole, while
intersection_area clipped it — two headline numbers on one panel describing
different populations. The count stays whole-feature, which is what "objecten"
means to an operator, but now reports how many the edge cuts and is marked an
estimate when it does. The area_weighted_sum branch reuses that same count
instead of issuing its own near-identical query.
Partitioned selection de-duplicated the count on source_feature_id but
returned the raw rows, so a building on a municipal boundary was counted once
and drawn twice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The greedy IoU matcher gave a reference to whichever candidate was offered
first. Row order decided that, and every detection in a run shares one
transaction timestamp, so ordering by created_at left the assignment
undefined: the same QA run over the same data produced different mean IoU,
and the geometry shown to a reviewer as a false positive could be the better
of two detections. Candidates are now ranked by confidence with feature
identity as tiebreaker, which is also the COCO/PASCAL rule.
A single precision/recall/F1 triple describes one operating point, so two
models cannot be compared from it: a conservatively calibrated model looks
worse at a low confidence cut and better at a high one without detecting
anything differently. DetectionMetricsService adds the full curve, average
precision and the threshold where F1 actually peaks.
Also:
- report the population the metrics were computed over, so
matches + false_positives equals candidate_feature_count even under an
area filter; raw dataset totals move to the _raw fields;
- state whether candidates are axis-aligned boxes or footprint polygons.
A box can never reach IoU 1 against a rotated building, so the strict
score has a ceiling that has nothing to do with detection quality.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Jens