Reference — Data Types

Every model this component serves belongs to one of four task families. A family decides what the model is fed and what its output tensors mean, and it produces one normalized output: a family-neutral shape that the result body carries and that the decision rules evaluate.

There is no bundle-supplied code. A model whose head no family can interpret is refused when the bundle is staged, so an unreadable head becomes a staging error rather than a wrong answer.

Normalized output

The normalized output has one field per family plus rawShapes, which records the shape of each output tensor the session returned. A family populates its own field and leaves the rest empty.

Field	Populated by	Meaning
family	every family	classification, detection, segmentation, or anomaly.
classes	classification	Scored classes, highest first.
detections	detection	Boxes surviving suppression, highest score first.
segments	segmentation	Per-class pixel counts and regions.
anomaly	anomaly	One score against one threshold.
rawShapes	every family	Output tensor name to shape. Diagnostic; a decision rule never needs it.

Coordinates are always normalized to the source image, in [0, 1], as [x, y, w, h] with the origin at the top left. A box decoded on a letterboxed model canvas is reported against the picture the camera took, so a consumer needs nothing but the result to draw it.

classes

Field	Meaning
label	The class name from the manifest’s label set, or class-<index> when the manifest names a count rather than labels.
index	The class index the model produced.
score	The activated score. With activation: softmax the scores of the full class set sum to 1; with sigmoid each class is scored independently; with none the raw value is reported.

The list holds at most topK entries, further bounded by the manifest’s maxResultItems, and drops anything below scoreThreshold.

detections

Field	Meaning
label	The class name from the manifest’s label set.
index	The class index, after classIndexOffset is subtracted.
score	Objectness times class score for a grid head; the head’s own confidence for a decoded head.
box	[x, y, w, h] normalized to the source image and clipped to it.

Suppression is class-aware: a high-scoring bolt never suppresses a washer occupying the same pixels, because two different things can be in one place. The list holds at most maxDetections entries, further bounded by maxResultItems.

segments

A mapping of label to {pixels, fraction, bbox}. Masks are never published.

Field	Meaning
pixels	How many pixels of the class map the class claims.
fraction	That count divided by the size of the class map, so a rule can be written against the image rather than against a resolution.
bbox	The region those pixels occupy, normalized to the source image, or null when the class claims none.

In argmax mode every label gets an entry, including the ones with no pixels, so a rule such as “no defect pixels” evaluates on a clean image instead of failing to resolve its path. In threshold mode the entry is the one named by positiveLabel. minPixels drops classes below that count.

anomaly

Field	Meaning
score	The activated, normalized score. In [0, 1] when the manifest declares a normalization.
threshold	The manifest’s threshold, in the same units as score.
anomalous	Whether the score crossed the threshold in the direction direction names.
direction	higherIsAnomalous or lowerIsAnomalous.
summary	Present for a map head only.

The summary of a map head carries min, max, mean, aboveThresholdPixels, fraction, and bbox — the region of the crossing pixels, normalized to the source image, or null when none crossed. The map itself stays in the executor cell.

Decision rules

The manifest’s decisionRules produce the result’s decision. The rules are JSONPath expressions over the normalized output, so a rule can name any field any family produces.

{
  "pass": { "path": "$.classes[0].score", "op": ">=", "value": 0.9 },
  "confidence": "$.classes[0].score",
  "threshold": 0.9,
  "outcomeOnPass": "CLEAR",
  "outcomeOnFail": "HOLD",
  "failOnEmpty": false
}

Key	Meaning
pass	The expression that decides the outcome. Required.
confidence	A JSONPath resolving to the number reported as decision.confidence. Optional.
threshold	A number, or a JSONPath resolving to one, reported as decision.threshold. Optional.
outcomeOnPass	The outcome when pass holds. Defaults to CLEAR.
outcomeOnFail	The outcome when pass does not hold: HOLD or FAIL. Defaults to HOLD. CLEAR is refused.
failOnEmpty	Whether a path that matches nothing is a plain failure rather than a broken rule. Defaults to false.

Expressions

An expression is a leaf or a group.

{ "all": [ {"path": "$.detections[*].score", "op": ">=", "value": 0.5},
           {"path": "$.detections[*].label", "op": "!=", "value": "washer"} ] }

Form	Meaning
{"path": …, "op": …, "value": …}	A leaf comparison.
{"all": [ … ]}	Holds when every child holds.
{"any": [ … ]}	Holds when at least one child holds.

Groups nest to any depth.

Operators

Operator	Meaning
>=, >, <=, <	Numeric comparison. Both sides must be numbers; a flag or a string is a broken rule.
==, !=	Equality against the literal in value. Works for numbers, strings, and booleans.
exists	The path matches at least one value. Takes no value.
absent	The path matches nothing. Takes no value.
count>=	The number of matched values is at least value.

A leaf whose path matches several values is a claim about every match: $.detections[*].label != "washer" means “no detection is a washer”. exists, absent, and count>= describe the match set itself rather than its contents.

Outcomes

Outcome	Meaning
CLEAR	The rule evaluated and passed.
HOLD	The rule evaluated and failed, or could not be evaluated at all.
FAIL	The rule evaluated and failed, and the manifest classes that failure as a defect rather than a doubt.

A rule that cannot be evaluated yields HOLD, never CLEAR. A missing path, a malformed expression, a confidence that resolves to nothing or to a non-number, a threshold that resolves to nothing, an outcomeOnFail of CLEAR: all of them hold the image. decision.rule says which rule decided, so an operator can tell a failed image from a failed rule set:

decision.rule	Meaning
pass	The rules evaluated and passed.
pass.all[1]: $.classes[0].score >= 0.99	This leaf is the one that failed.
pass.any: none of 2 matched	An any group had no child that held.
UNEVALUABLE:confidence: path '$.nowhere' matched nothing	The rule set is broken, not the image.

decision.confidence and decision.threshold are null whenever the outcome came from an unevaluable rule, because a number that could not be read is not a number to report.