Hub

A request (or a city need) arrives. The factory decides whether it can build it - cheaply by keyword-matching the need to its building-block machines, or smartly by asking the LLM autopilot. If yes, it builds a pipeline from those machines and runs it.

On the belt, each station does one job and passes a typed payload (a <PipelinePayload> of <Slice> rows) to the next - a station can read any earlier station's slice by name. Output from an untrusted station (the LLM, web search, external services) is checked at a trust boundary - validated against an XSD or a ContentType - before it flows on; trusted deterministic transforms just flow. The finished product carries a manifest slip (what it is + the route it took).

flowchart TD
  Req[Request or city need] --> Decide{Build it?}
  Decide -->|cheap keyword match| Build[Build pipeline]
  Decide -->|smart LLM autopilot| Build
  Build --> S1[Station 1]
  S1 -->|typed payload| S2[Station 2]
  S2 -->|typed payload| S3[Station 3]
  S3 --> Out[Product plus manifest slip]
  S1 -.->|untrusted output| Check{Trust boundary}
  Check -->|untrusted - validate vs XSD or ContentType| Reject[Reject if bad]
  Check -.->|trusted - just flow on| S2

flowchart TD
  Observe[Observe - Tool.CityStatus] --> Decide[Decide - what to build next]
  Decide --> Act[Act - build it or escalate a gap]
  Act --> City[(The living city)]
  City --> Observe

flowchart TD
  Eval["Eval part (XML) - trigger, scenario, tier-tagged assertions"] --> Det["Deterministic tier - stub model, real served stack"]
  Eval --> Qual[Quality tier - the real LLM offline]
  Det -->|"grades the PIPELINE - parses, validates, lands"| Grade[AssertionEvaluator - one vocabulary]
  Qual -->|grades the real ANSWER| Grade
  Grade -->|text assertions| Raw[Raw emission]
  Grade -->|XML assertions| Row[Landed row]
  Grade -->|lands-in-list| Store[(Storage re-read)]
  Grade --> Verdict["EvalVerdict - Pass, Fail or Degraded"]
  Verdict --> Result[(EvalResult row - queryable)]