Structured Output
Observation data is only useful if it arrives in a format your pipeline can actually use. Sentinel Watch delivers structured, schema-compliant datasets — timestamped, geolocated, and formatted for direct ingestion without transformation work on your end.
What Every Observation Record Contains
Each record delivered by Sentinel Watch is a structured object containing the following fields, populated at the point of observation and verified through the quality pipeline before delivery.
Taxonomy Fields
The event type, classification attributes, and any sub-classifications defined in your taxonomy. These are the fields that directly feed your model. They are captured using a structured form built from your schema — no free-text interpretation, no post-hoc normalisation.
Timestamp
Recorded at the moment of observation, not at submission. ISO 8601 format, UTC. For programs where temporal precision matters — behavioural pattern analysis, time-of-day segmentation, sequential event detection — this ensures the timestamp reflects when the event actually occurred.
Geolocation
Latitude and longitude recorded at capture. For programs with geographic segmentation requirements, this enables filtering, spatial analysis, and region-specific model training without additional enrichment steps.
Observer Identifier
An anonymised observer ID that allows your team to audit consistency across observers, identify individual drift, or segment the dataset by observer if required for analysis. The identifier is pseudonymous by default — observer identity is not exposed in the delivered data.
Quality Review Status
Each record carries a review status flag indicating whether it passed automated validation only, or also passed human review. Records that were adjudicated during the review process carry an annotation flag. This lets your team apply different confidence weightings to records if your training pipeline supports it.
Program Metadata
Program identifier, taxonomy version, and deployment context — enabling your pipeline to correctly handle data from multiple program phases, evolving taxonomies, or parallel geographic deployments without manual tracking.
Delivery Formats
JSON
Default format for programmatic ingestion. Each observation is a structured JSON object. Batch deliveries are newline-delimited JSON (NDJSON) for efficient streaming. Schema is versioned and documented.
CSV
Flat file format for direct ingestion into data warehouses, labeling platforms, or spreadsheet-based review workflows. Column headers match your taxonomy field names. UTF-8 encoded.
Custom Schema
If your ML pipeline, labeling tool, or data warehouse requires a specific output schema — proprietary field names, nested structures, platform-specific formats — we can configure delivery to match. No transformation work required on your end.
Delivery Methods
Batch Delivery
Scheduled delivery of completed observation batches to your specified endpoint — SFTP, cloud storage bucket, or direct download. Delivery cadence is agreed at program setup. Each batch includes the quality report alongside the dataset.
Incremental Delivery
For ongoing programs, observations can be delivered incrementally as they pass the review pipeline — allowing your training pipeline to ingest new data continuously rather than waiting for a full batch. Each incremental delivery includes only new, previously undelivered records.
MCP API
Enterprise clients with a dedicated MCP connection at app.sentinel-watch.org can query and retrieve observation data programmatically. AI agents and automation pipelines can filter, retrieve, and act on observation data without manual data transfers or scheduled deliveries.
Taxonomy Versioning
When your classification schema evolves between program phases, every record carries the taxonomy version it was captured under. This means your pipeline can correctly handle mixed-version datasets — applying different field mappings or confidence weightings to records from different phases — without manual tracking or annotation.
Schema change documentation is provided at every version update, including field-level change logs and migration guidance for your pipeline team.
Want to See a Sample Output File?
We can walk you through a sample observation dataset in JSON and CSV format, show you how the schema maps to your pipeline, and discuss any custom output requirements before a program begins.