# Four-round Adaptive Evidence Budget

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "four-round-adaptive-evidence-2026-08-06",
  "problems": [
    {
      "problemId": "three-round-adaptive-cap-is-too-tight",
      "summary": "The three-round scenario cap leaves only two adaptive queries after initial discovery, which can exhaust the budget before one final bounded alternative check.",
      "status": "optimized_unverified",
      "issueKeys": ["ST-2863", "TCID-13001"],
      "sourceRefs": [
        "https://pingcap.feishu.cn/docx/SDHAdVVOXoCCfJxpduOcX2hKnAh#problem:three-round-cap"
      ],
      "signalTextIncludesAny": [
        "three adaptive evidence rounds are insufficient",
        "no remaining round for the final bounded evidence check"
      ],
      "observedThrough": "2026-08-06",
      "effectiveAt": "2026-08-06T20:57:22+08:00",
      "verifiedAt": null,
      "attemptedAction": "Increase the existing fixed per-scenario batch limit from three to four rounds without changing its provider, target, window, read-only, per-round query, or shared deadline boundaries.",
      "evidenceUrl": "https://pingcap.feishu.cn/docx/SDHAdVVOXoCCfJxpduOcX2hKnAh"
    }
  ]
}
-->

- Optimization ID: `four-round-adaptive-evidence-2026-08-06`
- Owner: auto-oncall-services
- Source reference: Clinic and Investibook comparison follow-up
- Implementation status: deployed to staging
- Effective status: `optimized_unverified`

## Bounded change

Raise the existing fixed scenario limit from three batch rounds to four. Each
round still selects at most six read-only queries and shares the original
scenario deadline. Provider, organization, target, canonical incident window,
serialization, and no-fallback boundaries are unchanged.

The cumulative handoff limits increase from 18 to 24 query objectives and
artifact references so the fourth round is not silently truncated. Round-4
paths are admitted by the response schema, backend parser, Clinic and OP
materializers, workspace cleanup, and cumulative handoff merge.

## Verification status

- 2026-08-06 full test suite passed: 676 tests, 0 failures. Focused tests
  admit round-4 Clinic and OP artifact paths, reject round 5, and verify that
  the parent prompt exposes round 4 of 4.
- Source PR [#675](https://github.com/tidbcloud/aisupport/pull/675) and
  app-delivery PR
  [#76485](https://github.com/tidbcloud/app-delivery/pull/76485) deployed image
  `7c9240c8864aece2523606484925c8d91d83d250` to staging. The Deployment was
  1/1 Ready with zero restarts, both health endpoints returned 200, poller and
  worker logs were healthy, and the job table contained no queued or running
  jobs.
- A post-change case replay is required before `optimized_verified`; deployment
  alone may advance this record only to `optimized_unverified`.

## Unresolved gaps

- No real case has yet required and completed a fourth round.

## Next validation action

Replay a naturally multi-hop case later if fourth-round effectiveness needs
verification.