Ground truth

Ground truth in Utopia is a precise data quality guarantee, not a slogan. An observation carries the label only if physical sensors captured it in the real world, multiple independent nodes corroborated it, and a cryptographic hash signed at capture proves it was never altered. That is what separates Utopia's output from inferred, modeled or synthetic spatial data. Data of that kind may look statistically reasonable, but no one can independently verify it.

What ground truth means in Utopia's context #

Three criteria define a ground truth observation:

1. Physical capture. A real sensor recorded it, at a real location, at a real time. Nothing in it comes from a model, and nothing is interpolated from neighboring data.
2. Multi node corroboration. A consistent observation of the same feature was made by at least one other independent node inside the same window of space and time. Observations backed by a single node come back with a low ground_truth_score and an explicit unconfirmed marking.
3. Cryptographic provenance. The originating node generated a provenance_hash that binds the observation to a specific device, timestamp and content, giving tamper evident proof of where it came from.

Fail any of the three and the observation is not hidden. It is returned with a reduced score or a flag. Utopia's approach is to surface low confidence observations along with the metadata you need to judge them, never to suppress them.

How ground_truth_score is calculated #

Think of the ground_truth_score as a confidence metric on a 0.0 to 1.0 scale, driven by how strongly independent nodes corroborate an observation.

Three things push the score up: additional independent nodes reporting consistent observations of the same feature; contributing nodes that use different sensor modalities, because agreement across modalities counts for more than agreement within one; and a narrower temporal spread between the contributing observations.

Three things pull it down: an observation with only one contributing node; contributing nodes that disagree on position, classification or velocity beyond the configured tolerance thresholds; and any contributing node carrying a degraded hardware quality rating.

The score is a normalized measure of corroboration strength, not a probability estimate. Where you set your threshold is a function of how much uncertainty your application can absorb.

Ground truth versus inferred and synthetic data #

The clearest way to define Utopia's ground truth is by contrast.

Inferred data comes from a model estimating a scene's state out of partial observations or historical patterns. When a system fills the gap between sensor captures by running a motion model over vehicle positions, its output is inferred, potentially very accurate but never checkable against a physical observation.

Synthetic data never touched a sensor at all: simulation or generative models produced it. That has real value for training and testing, but what it describes is a virtual world, not the physical one.

Neither kind ever gets blended into a Utopia ground truth response. Each feature the API returns either came from a physical sensor or is explicitly labeled through its data_source value: interpolated when a historical record has been gap filled, or estimated when a single node observation is still awaiting corroboration.

How cryptographic provenance proves authenticity #

A provenance_hash field appears on every observation in every Utopia API response. The originating edge node generates it using a hardware backed private key, and it encodes three inputs: the node's unique identifier, the time and location the device recorded at the moment of capture, and a SHA-256 hash of the observation's feature payload.

Verification works by submitting the provenance_hash together with the feature payload to the Utopia Provenance API, which checks the signature against the node's public key and reports the result. A payload modified at any point after capture, whether in transit, inside Utopia's own infrastructure or inside your systems, fails that check.

Real time versus historical ground truth #

Two query modes exist, and which one you use changes how ground_truth_score and provenance_hash should be read.

Example API response with ground truth fields #

Below is one feature taken from a real time spatial query. The ground_truth_score, the provenance_hash and the fields around them are what jointly constitute a verified ground truth observation.

{
  "feature_id": "feat_9b1e3c7d",
  "type": "Feature",
  "geometry": {
    "type": "Point",
    "coordinates": [-73.9857, 40.7484, 8.1]
  },
  "properties": {
    "classification": "person.pedestrian",
    "velocity_mps": 1.3,
    "heading_deg": 45.0,
    "captured_at": "2026-05-22T14:32:05.198Z",
    "ground_truth_score": 0.91,
    "provenance_hash": "sha256:c7f2a49d1e8b3056a7f2c49d1e8b3056c7f2a49d1e8b30",
    "node_count": 4,
    "sensor_modalities": ["camera", "depth"],
    "data_source": "ground_truth",
    "corroboration_window_ms": 312
  }
}