Capabilities & operations

Operations.

For: operators

Day-to-day, running Cardinal is light: start the ring, keep the cadence where you want it, and watch the audit chain stay healthy. The consequential operation is recovery — bringing everything back on a fresh machine — and because the source of truth lives in the cloud, that is a time cost, not a data cost.

Light everyday ops Recover in ~1–2 hours Time, not data, at risk

Everyday operations

Routine

  • Start the ring with bootstrap; start the Ticker
  • Set the cadence posture — on-demand or scheduled
  • Read the brief; act on what needs you
  • Keep the corpus synced

When something's off

  • A listener went quiet → respawn it and drain the inbox
  • A role is tired → rotate it (a recorded ceremony)
  • A tool server won't start → re-run its auth, re-probe
  • The chain alarm fired → the Steward investigates first

Recovery after a laptop loss

The corpus lives in cloud storage and the database lives in Azure, so a lost machine takes no engagement data with it. Recovery is the same install you did the first time, in the same order.

StepWhat you do
1 · Restore the corpusPull it down from cloud storage onto the new machine.
2 · Re-authenticateSign in again in the canonical order — the slow, interactive part.
3 · Smoke-testConfirm every connection is live before spawning anything.
4 · BootstrapRe-spawn the ring and start the Ticker.
5 · VerifyEach agent's first action writes an audit row; confirm the chain verifies.
Latency, not loss. Scouts keep a per-source cursor, so even the in-flight work resumes where it left off. The bottleneck in recovery is the interactive sign-ins, not any data restore.

Cost & telemetry

Cardinal meters its own spend. Cost is attributed per role and per skill so it's clear where the run rate goes, and a budget breaker can hold spend under a ceiling. The cost story is described here by shape — what drives cost and how it's controlled — rather than by figures, which are engagement-specific.

What drives cost

  • Model calls per role and per skill
  • Embeddings for the conversation substrate
  • A small, steady infrastructure floor

How it's controlled

  • Per-role and per-skill attribution
  • A daily spend breaker on the costly paths
  • Model selection matched to each role's needs

Reference drift

A recurring operational lesson: when a shared reference document disagrees with the live system, the live system is the source of truth. Counts, versions, and inventories drift; verify against the database before trusting a number, and re-verify before publishing one.

The install this recovers → · The guarantees it preserves →