Expand description
DETECT: what an analyzer says it FOUND — the word, the confidence, and the two shapes a finding comes in.
§It is a FAMILY, and it is the SECOND here that is domain-neutral
geometry arrived as one because a point, a rectangle and a quadrilateral are
one geometry read at three arities. This house arrives as one for a narrower reason: a
classification is a WORD and a NUMBER, and the only thing that varies between its members is
whether the finding also has a place. Three types, one subject.
Nothing here knows what a label is a label OF. An image classifier’s tag, a sound event, a
zero-shot query hit, a barcode’s symbology and a document region’s kind are five callers with one
(word, confidence) pair between them, and the vision-specific PAYLOADS that pair one of these
with a pose, a mask or a set of landmarks are NOT this module’s — they belong to the crate that
has the poses, consuming these. That split is what makes the house domain-neutral rather than a
vision library with the vision taken out: the types below would be the same types for a PDF’s
classified regions, which have no camera anywhere near them.
§THE MODULE IS CALLED detect, and that is a census verdict
The alternative was a classify-class name, and the word this house is built from decided it:
the types are Detection and LocatedDetection<T>, so a module named classify would put a
reader through one translation on every path they ever wrote. The word is also the RULING’s own —
the seal that sent this cluster here names it, and it names an assess house beside it for the
frame-level measurements that carry no label at all, so detect is one half of a pair rather
than a name chosen alone.
The producer this framework has met calls the pair a classification and the type a detection in one breath — a labelled, scored classification with no geometry of its own — so no name here could have satisfied both words. The one that matches the types is the one that landed.
§THE MEMBERS, in two clusters
| type | holds | for |
|---|---|---|
Label | a non-empty word | the vocabulary half of the two below it |
Detection | a label and a Score | a WHOLE-FRAME classification — no geometry |
LocatedDetection<T> | a detection and a BoundingBox<T> | a finding WITH a place — box REQUIRED |
TextDetection<T> | words, a score and a box | what a RECOGNISER read out of a region |
Symbology | a barcode standard, open | the vocabulary half of the row below |
BarcodeDetection<T> | a payload, a symbology, a score and a box | what a DECODER read |
DocumentSegment<T> | a Quad<T> and a score | a four-cornered region a detector found |
§THE PAYLOAD TRIO IS A SECOND CLUSTER, and what separates it is what a WORD IS
The first three are built out of a Label — a CLASS a model chose from its own vocabulary. The
last four carry a PAYLOAD instead: a recognised string and a decoded code are the picture’s own
content read back, and they name nothing. Putting either in a label field would call one thing
by the other’s name in every schema that published it, so the two clusters have the same SHAPE
and different subjects — and a shape is not a reason to share a type.
The division is visible at the doors. Label::new refuses the empty word, because a
classification of nothing is not one; TextDetection::new and BarcodeDetection::new refuse
nothing at all, because a recogniser that found a region and no characters in it has made an
observation and the region is what a caller is asking about.
Symbology is the one closed-ish VOCABULARY here, and it is open anyway. Twenty-four named
standards read out of a documentation page, plus an escape carrying the folded word — so it takes
the ENUM filter’s four membership operators where a payload takes the text family’s twenty. See
that type, where the roster, the fold and the wire scalar are argued.
DocumentSegment<T> is the only member that holds a Quad<T>, and it is
therefore the only one whose columns are ALL VALUE — see below, where the derived leaf the other
two inherit is accounted for.
There is no Option<BoundingBox<T>> anywhere in this house. One type with an optional box
would make a classification that never had a location and a located finding whose box went
missing the same value; two types make the second unrepresentable. That is the ruling this house
was built to state, and every face below inherits it — there is no null box in a column, none in
a filter and none on the wire.
§WHERE AN INVARIANT LIVES: on the leaf, and nowhere else
A rule about ONE WORD is Label::new’s — non-empty, and nothing about case. A rule about one
number is Confidence’s and NormCoord’s, each in its
own house. Neither composite here checks anything at all, and that is not laziness: a label
says nothing about what a confidence may be, and neither says anything about where a rectangle
may sit. Both constructors are total and const, which is what building out of validated parts
buys — see avanalyze’s own contract, whose
subject-detection constructor is infallible for exactly this reason.
§THE ENGINE OWNS NORMALISATION
Stated once here because it decides the shape of Label and therefore of everything above it:
the engine lowercases and trims image-classification labels but passes animal labels through
verbatim, and uses the literal "person" for human subjects — the label vocabulary is Apple’s,
not this crate’s. So Cat and cat are two labels, the type folds nothing, and a caller that
wants them to be one asks with the nocase half of a text filter’s affix slots. See Label,
where the division is argued at the constructor that keeps it.
§THE GATE: none, on the geometry house’s own precedent
Every vocabulary door in this crate is named for the DEPENDENCY it admits. This house admits
none: Label is text over the crate’s own carrier, Score is the crate’s
own confidence, and the geometry it composes is ungated beside it.
It needs no TIER either, and that is a measurement rather than a preference. The
#[ingraph(composed)] derive takes a CONCRETE struct — which is why
lang rides derive for a filter this house’s sibling declares — and
LocatedDetectionFilter<CF> is generic over its component’s filter, so it could not be
declared through that word whatever this module wanted. Declaring ONE of the two and
hand-writing the other would have bought a derive gate for half a house. Both are hand-written
on TimeRangedFilter’s road, and a build that declines the
declaration tier keeps every one of them.
Each FACE beneath the family carries its own cross-gate, as every family here does: sqlx for
the binds, decodes and lowerings, mongo for the documents, async-graphql for the wire.
§The three layers, and the crossings between them
domain crossing storage
Label identity text
Symbology CANONICALISING text
Detection per half (text, f32)
LocatedDetection<T> per half (text, f32) + the box's FIVE
TextDetection<T> per half (text, f32) + the box's FIVE
BarcodeDetection<T> per half (text, text, f32) + the box's FIVE
DocumentSegment<T> per half the quad's EIGHT + f32Every crossing is a HALF’s, applied per column, and ONE of them canonicalises: Symbology
folds ASCII case and three separators at its door, so DATA-MATRIX and data_matrix are one
value and a stored slug reads back as the named member. That is a vocabulary’s business and not a
payload’s — Label, a recognised string and a decoded payload each rewrite NOTHING, so a value
this crate wrote reads back byte for byte at every door.
The located finding INHERITS a derived leaf, and it is the geometry house’s.
BoundingBox<T> stores an area beside the four numbers it holds, so a
subject: LocatedDetection<NormCoord> spans SEVEN columns of which SIX are the value. This house
computes nothing — the fifth box column is the box’s own bind — and what it owes is the
CONSEQUENCE: a holder of a citizen with a derived leaf owes an order of its own, because a wide
term does not know that one of the columns beneath it is an index. See
LocatedDetection<T>’s ORDER, where six levels are named over seven leaves.
§What a DECODE re-proves, and it is everything
A read goes through each half’s own constructor, so an empty stored label is refused at the
column that held it and a confidence outside [0.0, 1.0] at its own. Unlike the geometry house
one door over, there is no cross-column rule to lose: a detection’s two columns are
independent and a located detection’s six are its parts’, so the composites promise nothing at
construction that a decode cannot re-establish. The recorded gap in the sibling family — a sum
rule no Assemble can re-check — has no analogue here.
§WHAT IS REFUSED, BY NAME
No whole-value operator, at either arity. No eq, ne, in or ordering over a whole
detection: the leaf tuple IS the conjunction of the facets and is reachable as one, a whole-value
ne over a composite is a disjunction, a per-column in is the cartesian product, and an order
composed of a word’s and a float’s is an order over the storage that would read like a ranking.
See DetectionFilter, where each clause is argued.
No FIXED-WIDTH cursor, and this house frames its members instead. The composites’ keyset
positions are written with write_member and read with
read_member — the crate’s one authority for several values in
one buffer, four bytes of big-endian length a member — where the geometry family divides its
payload by a constant. The division is unavailable here: a label is variable-width and it is
FIRST in segment order, so there is no constant split point from either end. See Detection’s
own cursor row, where the choice is argued; what four bytes a member buys is a read that refuses
a frame this type did not write.
No Vec<Label>, which is Utf8Bytes’ own position inherited: the
crate’s text carrier is not a CollectionElement on any backend, and
neither are the lang subtags. A list of words is a
Vec<Detection> in every declaration this framework has met — a bare tag has no confidence to
rank it by — and THAT is admitted, along with Vec<LocatedDetection<T>>.
No MONGO face on the four citizens that hold GEOMETRY, and it is a RECORDED GAP rather than a
refusal.
Label, Detection and Symbology all have their document forms, but a located finding,
a reading, a decoded code or a segment would need BoundingBox<T> or
Quad<T> to have a MongoCarrier, and the geometry house
publishes none for any of its three holders. Writing one here would put the geometry family’s
BSON shape in the detect family’s file, decided by the house that consumes it rather than the
house that owns it. So the door is named and left at the type that has to open it.
§THE DERIVED LEAF FOUR OF THESE INHERIT, and the sequence each of them owes
BoundingBox<T> stores an area beside the four numbers it holds — a
DERIVED leaf, computed by the write and read by no decode. Nothing in this house computes
anything; what it inherits is an OBLIGATION, and it is worth stating once for all four:
A holder of a citizen with a derived leaf owes a sequence of its own.
A composite that states no ORDER keeps a WIDE term whose keyset row
value is its whole width — so a finding, a reading or a decoded code wearing the non-reordering
word would put …_bbox_area inside a pagination cursor, which is the 2026-08-29 law’s own
prohibition. So LocatedDetection<T>, TextDetection<T> and BarcodeDetection<T> each
state a sequence and publish a leaf vocabulary, and each of those sequences is the DECLARATION
order minus the box’s derived leaf: the point is the EXPANSION, not a claim about how findings
sort.
DocumentSegment<T> states none and needs none, and the contrast is what makes the rule
legible: a quad is four positions and a position is two components, so nothing beneath it
derives anything and every one of its nine columns holds the value.
Structs§
- Barcode
Detection - A DECODED CODE, the standard it was read under, and where on the frame it was.
- Barcode
Detection Filter - The filter that guards a
BarcodeDetection<T>column: the payload, the standard, the confidence and the region. - Detection
- A CLASSIFICATION — what something was decided to be, and how sure the decider was.
- Detection
Filter - The filter that guards a
Detectioncolumn: the WORD’s own filter, and the CONFIDENCE beside it. - Document
Segment - A FOUR-CORNERED REGION, with the detector’s confidence in it.
- Document
Segment Filter - The filter that guards a
DocumentSegment<T>column: the four corners, and the confidence. - Empty
Label - What a label refuses a value with — one sentence, at the four doors that mint one from the outside: a caller’s constructor, a stored column, a cursor a client handed back, and a wire operand.
- Label
- A CLASSIFICATION’S WORD — what a detector, a classifier or a tagger says a thing IS.
- Located
Detection - A CLASSIFICATION AND ITS PLACE — what was found, how sure the finder was, and WHERE.
- Located
Detection Filter - The filter that guards a
LocatedDetection<T>column: the CLASSIFICATION axis flattened, and the GEOMETRY beside it. - SqlBarcode
Detection Comparison sqlx BarcodeDetectionFilter’s storage-side twin — four comparisons, and the two text-shaped ones reached through two different kinds.- SqlDetection
Comparison sqlx DetectionFilter’s storage-side twin — the label’s comparison and the confidence’s, and nothing of its own.- SqlDocument
Segment Comparison sqlx DocumentSegmentFilter’s storage-side twin — the quad’s whole comparison and the confidence’s.- SqlLocated
Detection Comparison sqlx LocatedDetectionFilter’s storage-side twin — the classification’s whole comparison, and the box’s beside it.- SqlText
Detection Comparison sqlx TextDetectionFilter’s storage-side twin — the words’ comparison, the confidence’s, and the region’s.- Text
Detection - WORDS THAT WERE READ, and the region they were read from.
- Text
Detection Filter - The filter that guards a
TextDetection<T>column: the WORDS, the confidence, and the region.
Enums§
- Barcode
Detection Leaf - The VALUE one of a decoded code’s seven levels compares.
- Barcode
Payload Kind - THE KIND OF THING a barcode’s decoded payload names — ten named kinds, plus an escape for a word this vocabulary does not name.
- Label
Wire - How a label crosses when the framework’s own text filter holds it — the reading that gives
TextFilter<Label, LabelWire>a face. - Located
Detection Leaf - The VALUE one of a located finding’s six levels compares — the per-leaf vocabulary this citizen publishes beside its sequence.
- Symbology
- THE BARCODE STANDARD a payload was encoded in — twenty-four named symbologies, plus an escape for a word this vocabulary does not name.
- Text
Detection Leaf - The VALUE one of a reading’s six levels compares.