@@ -30,6 +30,10 @@ Verify performance with comparable evidence rather than isolated job duration.
    Use only naturally occurring project runs; never create synthetic,
    benchmark, soak, canary, rerun, or empty-commit traffic to fill a cohort.
 3. Decompose median, p90, and p95 latency into queue, provision, setup, execute, upload/deploy, and teardown stages.
+   Use `queue.assigned`/`queue.running` spans for queue lifecycle and
+   `provider.create_instance`/`provider.delete_instance` for provisioning and
+   teardown. Join by `runner_name`; retain `queue_job_uuid`, workflow run and
+   `incus_member` so member pressure is not attributed to the services host.
 4. Compare throughput, utilization, success, retry, infrastructure-failure, flaky-failure, and missing-telemetry rates.
 5. Identify critical-path changes and resource contention; keep priority classes separate.
 6. Reject conclusions with insufficient samples or incomparable cohorts and label them `NOT_PROVEN`.
 
@@ -37,7 +37,17 @@ Explain fleet health and bottlenecks from correlated, time-bounded evidence.
    correlation fields. Distinguish the bounded pre-reconciliation account-only
    window from intent records that remain unbound after authoritative queued
    and already-running reconciliation.
-7. Separate confirmed faults, saturation, waste, and `NOT_PROVEN` gaps.
+   Treat `gha_fleet_queue_missing_runner_request_id` as incomplete
+   pre-execution correlation only. A running direct-JIT job with UUID,
+   workflow-run, numeric GitHub-runner and runner-name identity belongs to
+   `gha_fleet_queue_direct_jit_without_runner_request_id`; GitHub emits no
+   AcquireJobs request ID for that path.
+7. In OpenObserve, query job-start logs by `msg = 'job correlation accepted'`.
+   Join `github_*`, `runner_name` and `instance_name` to `fleet_traces` using
+   trace search type `traces`; queue spans expose `queue_job_uuid`, while
+   provider spans expose `incus_member`. Do not use the provider process
+   resource host as the compute placement member.
+8. Separate confirmed faults, saturation, waste, and `NOT_PROVEN` gaps.
 
 ## Safety
 
 
@@ -26,12 +26,26 @@ Trace one run from event receipt to terminal result without changing state.
 ## Workflow
 
 1. Capture repository, run, attempt, job, event, requested labels, commit, and timestamps.
-2. Follow the same identifiers through admission, priority, assignment, provisioning, runner registration, setup, execution, upload, and teardown.
+2. Follow the same identifiers through admission, priority, assignment,
+   provisioning, runner registration, setup, execution, upload, and teardown.
+   Start from the native job-start log fields (`github_repository`,
+   `github_workflow_run_id`, `github_run_attempt`, `github_job_name`,
+   `github_workflow_ref`, `github_commit_sha`, `runner_name`, `instance_name`).
+   Join traces by `runner_name`: `queue.*` supplies `queue_job_uuid`, and
+   provider create/delete supplies `incus_member`.
 3. Locate the first abnormal stage; distinguish queue pressure, missing capability, infrastructure failure, deterministic test failure, flaky failure, timeout, and observability loss.
 4. Compare with adjacent successful runs using the same workflow and toolchain.
 5. If a transient infrastructure error is proven, recommend a bounded rerun; do not execute it here.
 6. Preserve running jobs and label unsupported claims `NOT_PROVEN`.
 
+## Direct-JIT identity
+
+Do not require `github_runner_request_id` from a running direct-JIT job. It is
+an AcquireJobs identity GitHub does not emit on that path. Require queue UUID,
+workflow run, numeric GitHub runner, runner name and instance/member correlation
+instead. A missing request ID remains actionable only before the execution path
+has converged or when those authoritative identities are also absent.
+
 ## Output
 
 Return the causal timeline, first failing stage, supporting links or private evidence locations, impact, and safest recovery path.