One reporting plane, several honest kinds of evidence
Applications can emit low-cardinality Web Foundations request records. Caddy can retain separately filtered edge evidence. Tend reports bounded deployment phases and outcomes. Agents collect selected Linux metrics and OTLP signals. Observatory assigns organization, project, environment, and service scope from the enrolled source—not from an untrusted payload.
Those signals can share a timeline and correlation IDs without pretending they are interchangeable. A release marker can sit beside request rate, latency, errors, memory, and traces while the application and Tend remain authoritative for their own operational state.
Time series are a view of preserved evidence
The agent batches and spools outbound observations. The server commits immutable raw segments before acknowledging projection work. Common indexed views and rollups make ordinary dashboards fast; old or unusual questions can be rebuilt from retained evidence rather than forcing every possible field into an eager database schema.
Saved text queries and the visual builder compile to the same typed AST. Every query has time, row, scanned-byte, memory, and execution limits. Time-series charts retain an accessible table, and pinned dashboards do no work merely because nobody is watching them.
Collection policy belongs to the operator
Safe defaults do not erase useful evidence. IP addresses, queries, referrers, user agents, and session correlation can be collected when an application's privacy policy and operator decision require them; they remain explicitly classified and separately authorized. Unknown fields default to sensitive and unindexed rather than silently becoming public dashboard material.
The agent has no Docker socket, root requirement, arbitrary shell plugin, or inbound collector port. Its configured sources stay locally authoritative, and an Observatory outage cannot block the applications or Tend deployments it observes.
Dogfooding the complete path
Gamertan now emits Web Foundations request evidence for Observatory to normalize and query. Tend already provides the separate deployment stream. That gives this site a real request-to-release timeline and gives Observatory a first-party workload whose privacy, latency, retention, and failure boundaries I can keep improving in public operation.