# TiCDC Architecture Evidence Boundary

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "ticdc-architecture-evidence-boundary-2026-08-28",
  "problems": [
    {
      "problemId": "ticdc-runtime-chain-rejected-without-explicit-architecture-field",
      "summary": "Auto-oncall product knowledge required an explicit architecture field or startup marker, while the Report Judge did not receive any TiCDC architecture evidence boundary and could reject a coherent new-architecture runtime chain.",
      "status": "optimized_unverified",
      "applicability": "TiCDC architecture diagnosis or AI-versus-human evaluation where direct control-plane architecture metadata is absent but exact-changefeed runtime evidence is available.",
      "issueKeys": ["TCID-13093"],
      "sourceRefs": [
        "https://tidb.atlassian.net/browse/TCID-13093#problem:ticdc-runtime-chain-rejected-without-explicit-architecture-field"
      ],
      "signalTextIncludesAny": [
        "controlPlaneArch",
        "new-architecture runtime path",
        "explicit architecture field",
        "Maintainer",
        "Event Service"
      ],
      "observedThrough": "2026-08-28",
      "effectiveAt": null,
      "verifiedAt": null,
      "attemptedAction": "Define one three-tier TiCDC architecture evidence boundary in repository product knowledge and project its controlled Judge guidance into relevant Report and single-ticket comparisons.",
      "evidenceUrl": "https://tidb.atlassian.net/browse/TCID-13093"
    }
  ]
}
-->

- Optimization ID: `ticdc-architecture-evidence-boundary-2026-08-28`
- Owner: TiCDC repository product knowledge; Report Judge receives a bounded
  projection from that same source
- Source case: `TCID-13093`
- Implementation status: local candidate
- Effective status: `optimized_unverified`

## Bounded problem

The previous topic admitted only an explicit architecture field or aligned
startup/deployment marker as proof. It therefore treated an exact-target,
exact-window chain across new-architecture-only runtime stages as unknown. The
Report and single-ticket Judge independently lacked this product boundary, so a
diagnostically valid runtime-path conclusion could still be scored as too
strong merely because `controlPlaneArch` was absent.

## Evidence contract

- Direct proof: explicit control-plane architecture metadata or an aligned
  startup/deployment marker.
- Equivalent runtime-path proof: coherent same-changefeed, same-window activity
  across Maintainer, dispatcher, Event Service, and where observable Event
  Store or Log Service. The answer must disclose that direct metadata or a
  startup marker was absent.
- Insufficient alone: TiCDC version, `arch=dedicated`, `deploy_type`, topology,
  Dataflow identity, metric-family existence, or one owner/dispatcher series.
- Classic compatibility metrics remain counterevidence; they do not
  automatically veto a complete aligned new-architecture runtime chain.

## Verification criterion

After deployment, rerun `TCID-13093` in `mode=read`. The Scenario Diagnostic
Agent must preserve the direct-field absence while recognizing the aligned
runtime chain as execution through the new-architecture path. Run the same
frozen answer through the single-ticket Judge; the Judge must apply the same
three tiers and must not cite direct-field absence alone as an AI gap. Only
then run the authorized write replay and read back the stable `public=false`
Jira comment before marking this problem `optimized_verified`.