Analyzer SDK

The export bundle

There are two export shapes, and which one you get depends on whether the detector has weights.

Trained modelZIP

model-export.json at the root, plus weights.onnx beside it. The JSON names the weights file in its weightsPath.

AlgorithmJSON

The same model-export.json document on its own, with weightsPath set to null. Not a ZIP, because there is nothing to bundle alongside it.

Both shapes are the same document type. A learned export is that specification plus a referenced weights file, so one reader handles both.

{
  "specVersion": 1,
  "modelId": "mdl-7f3c1a",
  "name": "Presence (living room)",
  "version": "1.0.0",
  "algorithmId": "algo-presence",
  "kind": "deterministic",
  "architecture": null,
  "filterIds": ["denoising", "bandpass"],
  "filterImplVersions": [
    ["denoising", "1.0.0"],
    ["bandpass", "1.2.0"]
  ],
  "baselineSessionId": null,
  "weightsPath": null,
  "accuracy": null,
  "f1": null,
  "exportedAt": "2026-08-27T09:14:02Z",
  "recipe": {
    "specVersion": 2,
    "algorithmId": "algo-presence",
    "kind": "deterministic",
    "pipeline": [{ "op": "temporal_presence_optimized" }],
    "output": { "format": "boolean", "deriveFrom": "present" }
  }
}

Filters and reproducibility

filterIds names the preprocessing stages the detector was built against. filterImplVersions gives the implementation version of each one.

Two implementations of the same named filter can produce different outputs from a byte-identical specification: the ids match, the names match, and the answers differ. A learned model is pinned by the content hash over its weights; a closed-form detector has no equivalent, so its stage versions are its version.

The recipe document

A recipe is a declarative pipeline: an ordered list of operations over the CSI, plus a statement of how the result is shaped. When a bundle carries one, the SDK executes the recipe — it does not look up a compiled algorithm by name. That is what makes an exported detector portable rather than a reference to something on the other side.

specVersionnumber

Format version of the recipe; the current version is 2. A reader rejects any version above the one it knows. Version 1 documents are accepted and run unchanged.

algorithmIdstring

Which detector this is. Must match the algorithmId on the enclosing export document.

kindstring

deterministic, calibrated, or learned. Checked against the pipeline: a recipe declaring learned must contain an inference step, and one containing an inference step must declare learned. A document that runs weights while calling itself closed-form is not reproducible from the document alone.

pipelinearray

One or more steps, executed in order. An empty pipeline is rejected. Each step names an op and may carry a params object; a parameter the op does not accept is an error rather than something ignored.

outputobject

How the pipeline's result is exposed. See below.

The document is validated as a whole before anything runs, and that includes the shape of any fitted numbers inside it. A classifier whose mean rows do not match its class list is rejected at load rather than scoring against the wrong column on somebody's recording.

Pipeline operations

A step names an op. Anything not on this list is rejected — an unknown op is a document a reader cannot honour, and skipping it would produce an answer from a pipeline that never ran.

temporal_presencedetect

Presence over a window from temporal variation.

temporal_presence_optimizeddetect

Presence with tuned thresholds. The default presence step.

temporal_headcountdetect

How many occupants, given presence. Meaningless on its own — presence has to have been established first.

temporal_stationary_holdstabilise

Holds a decision across intervals where an occupant is still. Without it, someone sitting motionless reads as an empty room.

median_smoothstabilise

Median filter over the decision series, removing single-window flips.

capture_breathing_windowsmeasure

Per-window breathing rate and a signal-quality figure. Produces a time series rather than a decision.

capability_gategate

Suppresses downstream steps when the recording does not have what they need.

diag_mahalanobis_countclassify

Diagonal-Mahalanobis classification over named features. Requires a classifier on the same step; no other op accepts one.

onnx_inferinfer

Runs the bundle's weights over the capture. Takes artifact (the weights file inside the bundle) and emits (count or presence). A recipe containing this step must declare kind: "learned". Added in spec version 2.

require_presencegate

Zeroes the count wherever the presence series says the space is empty. Takes target: "count". Added in spec version 2.

gate_breathinggate

Suppresses breathing measurements for windows that are unoccupied or too active for a breathing rate to mean anything. Takes by: "presence". Added in spec version 2.

"pipeline": [
  { "op": "temporal_presence" },
  { "op": "temporal_headcount" },
  { "op": "temporal_stationary_hold" },
  { "op": "median_smooth", "window": 5 }
]

Step parameters

A step's settings go in a params object keyed by name. Each op declares which parameters it accepts, and a key the op does not take is rejected when the document is loaded — not ignored.

"pipeline": [
  { "op": "temporal_presence" },
  { "op": "temporal_headcount" },
  { "op": "require_presence", "params": { "target": "count" } },
  { "op": "median_smooth", "params": { "window": 9 } }
]

Fitted numbers as data

A calibrated detector's parameters live in the document, not in code. diag_mahalanobis_count carries per-class feature means in raw units and one pooled within-class variance per feature.

{
  "op": "diag_mahalanobis_count",
  "classifier": {
    "classes": [0, 1, 2],
    "features": ["wb_tap_power_max", "wb_tap_power_sum"],
    "means": [
      [141.0, 402.0],
      [1232.0, 3110.0],
      [2084.0, 5233.0]
    ],
    "variances": [88.5, 240.1]
  }
}

The numbers are in raw feature units and there is no scaler to apply. Diagonal Mahalanobis over raw features with pooled variances is the same rule as over z-scored features, so a reader scores a capture directly against the means and variances in the document.

Composing two detectors

A pipeline can mix a closed-form detector with a learned one. The composition happens at the decision level: one detector's answer becomes an input to the other's, with no shared feature vector anywhere.

{
  "specVersion": 2,
  "algorithmId": "algo-presence-gated-count",
  "kind": "learned",
  "pipeline": [
    { "op": "temporal_presence" },
    { "op": "onnx_infer", "params": { "artifact": "weights.onnx", "emits": "count" } },
    { "op": "require_presence", "params": { "target": "count" } },
    { "op": "median_smooth", "params": { "window": 5 } }
  ],
  "output": { "format": "count", "deriveFrom": "present" }
}

Presence gates the count, and not the reverse. A headcount read from an interval whose presence detector says the room is empty is reading noise.

Output

output states the shape of the answer, which is what lets one consumer render every detector without knowing which one it is holding.

formatstring

boolean, count, or timeseries.

deriveFromstring

For boolean and count: present or count_gt_zero. For timeseries: breathing_windows.

channelsarray

Required for timeseries, and rejected for the other two. Each channel is a label and a field; the accepted fields are bpm and snr.

"output": {
  "format": "timeseries",
  "deriveFrom": "breathing_windows",
  "channels": [
    { "label": "breaths/min", "field": "bpm" },
    { "label": "signal quality", "field": "snr" }
  ]
}

Exporting and importing

Export is a download from the detector's own page: a trained model from the model detail panel, a catalog algorithm from its card. Import is a single upload that accepts either shape and figures out which it received.

An import is not a restore. Several fields on the bundle name rows in the organization that produced it, and carrying them across would be meaningless at best:

• A fresh id is minted. The bundle's modelId is recorded as the source and never reused, so an import can never overwrite an existing detector — including one in an organization the importer does not belong to.
• Visibility is forced private. A bundle that was public where it was made does not become public where it lands; sharing it is a separate, deliberate act.
• Session and training-pack references are dropped. They name recordings in another organization, which the importer cannot see and must not be told about.
• The recipe is kept. It is the executable part — an import that dropped it would produce a row that looks like a detector and cannot answer anything.