# Residual-alternative causal strength

<!-- auto-oncall-optimization-record
{
  "schemaVersion": "auto-oncall-optimization-record-v1",
  "optimizationId": "residual-alternative-causal-strength-2026-08-14",
  "problems": [
    {
      "problemId": "unobserved-residual-alternative-promoted-by-elimination",
      "summary": "After checked causes are weakened or not observed, an unobserved residual explanation can be promoted from a remaining investigation direction to the most likely cause without its own positive entity/time-aligned evidence.",
      "status": "optimized_verified",
      "applicability": "Provider-backed diagnostic cases where one or more checked causal candidates are weakened and a materially different residual explanation remains unobserved.",
      "issueKeys": ["APID-29333"],
      "sourceRefs": [
        "https://tidb.atlassian.net/browse/APID-29333#problem:diagnostic_causal_strength:residual-alternative-promotion"
      ],
      "signalTextIncludesAny": [
        "unobserved residual promoted to most likely",
        "remaining direction called the leading cause without positive evidence",
        "eliminating checked candidates treated as evidence for another cause",
        "background activity ranked without entity time aligned evidence"
      ],
      "observedThrough": "2026-08-14",
      "effectiveAt": "2026-08-14T14:46:43Z",
      "verifiedAt": "2026-08-14T15:38:57Z",
      "attemptedAction": "Require the Scenario Diagnostic Agent to keep an unobserved residual alternative viable but unverified, and require the Scenario Router to rank it as leading or most likely only from its own positive entity/time-aligned mechanism evidence or an observed cross-scenario dependency. Preserve bounded leading conclusions when positive aligned evidence exists even if a weaker alternative remains viable.",
      "evidenceUrl": "https://tidb.atlassian.net/browse/APID-29333?focusedCommentId=23528578"
    }
  ]
}
-->

- Optimization ID: `residual-alternative-causal-strength-2026-08-14`
- Owner: Scenario Diagnostic Agent and Scenario Router
- Source reference: `APID-29333`
- Implementation status: deployed to staging in image `9ebd2d3e8160d5321c62749043a32b9eb471396f`
- Effective status: `optimized_verified`

## Bounded problem

The replay run `866a02d0-b8e7-4a96-8f11-b50254dad59f` weakened the checked
frontend-load and PD-role explanations, then described unobserved background
activity as the current most likely direction. The evidence only supported it
as a remaining direction to investigate. Eliminating other candidates did not
provide positive evidence for that explanation.

This is a causal-strength decision owned first by the Scenario Diagnostic
Agent for its local alternative and then by the Scenario Router for the one
ticket-level ranking. It is not a response-format or Final Response problem.

## Candidate change

- The Scenario Diagnostic Agent lowers only the candidates addressed by
  negative or absent observations. It keeps a residual explanation viable but
  unverified unless that explanation has its own entity/time-aligned mechanism
  evidence.
- The Scenario Router requires the same positive support, or an observed
  cross-scenario dependency, before using leading or most-likely wording.
- Positive aligned evidence may still support a bounded leading conclusion
  without first eliminating every weaker alternative.
- Final Response remains wording-only and receives no causal-ranking rule.

## Verification evidence

- Focused prompt-contract tests cover the negative residual case, the positive
  aligned-evidence preservation case, and the Final Response non-owner
  boundary.
- PR `tidbcloud/aisupport#924` and app-delivery PR
  `tidbcloud/app-delivery#77344` are merged. Staging Deployment
  `auto-oncall-services` became Ready at `2026-08-14T14:46:43Z` with image
  `9ebd2d3e8160d5321c62749043a32b9eb471396f`; `/healthz` and `/readyz` both
  returned HTTP 200.
- Frozen-cutoff read run `a99ae719-a755-4dce-bcd7-1a5f79253ddd` kept SQL and
  background activity as time-related candidates and explicitly declined to
  call them the cause without node/time alignment.
- Write run `814ba2de-57f7-45c6-ac62-4adacb95093f` completed provider
  acquisition, received Judge verdict `ai_better`, persisted DB run `150049`,
  evaluation, and Feishu updates without errors, and updated existing Jira
  internal comment `23528578`. Direct readback verified `public=false` and the
  bounded causal wording.
- Preservation runs `355f6651-008c-48ac-b351-4e6b596c894a`,
  `0ac09551-dad1-4cac-ab69-a71226377592`,
  `7bd51ab4-c2e8-4fc1-ad3c-f63f45d9631f`, and
  `07c62d47-5c11-42dc-96e6-b35ef8163ad0` retained bounded alternatives,
  ordinary refinement, and positive-evidence leading conclusions. The problem
  is `optimized_verified`.

## Next validation action

Monitor newer diagnostic evaluations for the same residual-promotion signal;
no further action is required for the bounded APID-29333 problem.