Read-only developer reference

The honesty contract

trusted_pct and delivered_pct answer different questions. Neither can replace the other.

• trusted_pct: of the samples that arrived, how many were believable (quality 0–2)?
• delivered_pct: against each metric's catalogued cadence, how many expected samples arrived?

A producer emitting one sample every ten minutes for a 60-second metric can be 100% trusted and only 10% delivered. Showing trust alone would hide 90% loss.

Denominator and roster

The expected count is cadence-derived per metric. The response states roster_basis: series reporting in this window or the preceding equal-length window. Series below 90% delivered appear in silent_series; a missing series is evidence, not zero.

Provisional windows

Offline edge nodes can backfill. The late-data horizons are 24 hours for raw/5-minute, 3 days for hourly and 7 days for daily aggregates. Inside that horizon, is_provisional is true and provisional_from states the boundary.

Do not draw across outages

Empty buckets are omitted from the points array. A client that connects the received points will draw a straight line across an outage. Build the expected bucket grid from the resolution, insert gaps as null values, and render a line break. Read delivered_pct and silent_series beside the chart.