成果の指標 — Yardドキュメント

Technical reference synchronized with Yard’s source. File names and development commands refer to the native app repository.

NativeRunInsights exposes deliveries alongside the existing attempt, run, model and deployment summaries. NativeDeliveryMetrics projects the canonical evidence journal; it does not infer a delivery episode from legacy attempts, a board column, or an agent's completion claim.

Episodes and cohorts

delivery.started starts an episode identified by data.delivery_id. The identity is scoped to the project and is retained across attempts until delivery.accepted or delivery.abandoned. Reopening a delivered task creates another episode. Changing the captured contract ends the previous episode and creates another; a task number does not combine all its historical work forever.

The reporting cohort contains episodes whose terminal event occurred in the selected period. Costs, attempts and evidence include that episode's full earlier history. Active episodes with activity in the period appear separately as incomplete. Events after the reporting end are excluded.

An event supplies data.delivery_id directly or inherits attribution through its context.attempt_id from task.started. Unattributed agent calls remain visible in the coverage fields and are not assigned to a convenient task.

Definitions

Field	Meaning
`accepted`	Explicitly delivered episodes with a confirmed merge SHA. Observation mode can deliver with incomplete acceptance evidence. The UI labels this count Delivered.
`verified_accepted`	Delivered episodes whose final integration evaluation passed all captured mandatory criteria, with matching reviewed head and contract hash. A previous QA report cannot establish this.
`first_pass_rate`	Delivered episodes with one successful QA and Review result, no failed stage and at most one task attempt. Missing stage evidence remains unknown.
`lead_time_ms`	Observed Ready timestamp to delivery. An observed Ready timestamp must be no later than dispatch.
`dispatch_to_acceptance_ms`	First episode dispatch to delivery, including retries and waits. The UI labels it Dispatch to delivery when Ready coverage is incomplete.
`active_wall_ms`	Union of observed agent and evaluator execution intervals. Parallel work counts once.
`active_compute_ms`	Sum of those execution intervals. Parallel work counts for each invocation.
`queue_ms`	Observed Ready to first dispatch. Missing Ready evidence is not a zero queue.
`policy_wait_ms`	Completed intervals explicitly recorded for policy approvals.
`unplanned_wait_ms`	Completed intervals explicitly recorded for clarification, correction or rescue.
`known_cost_per_accepted`	Known costs of all attempts and evaluators attributable to the delivered cohort divided by its delivered episodes.
`known_cost_per_verified_accepted`	Same attribution restricted to deliveries with verified acceptance. The UI uses this metric.
`accepted_cost_coverage`, `verified_accepted_cost_coverage`	Calls with reported cost divided by all attributed calls for that cohort. Estimated costs remain separate from reported costs.
`abandoned_known_cost_usd`, `incomplete_known_cost_usd`	Known spend on abandoned and active episodes, separately from delivered work.
`unverified_delivery_known_cost_usd`	Known spend on delivered work without verified acceptance.
`requirements_pass_rate`	Passed required criteria divided by all required criteria in each episode's latest evaluation. Failed and unverified criteria remain explicit.
`autonomy_rate`	Delivered episodes without unplanned human intervention, among episodes with explicitly complete intervention observation. Policy approval is separate.
`observed_regressions`	Explicit bug/regression feedback linked to the delivered merge SHA.

The absence of feedback does not demonstrate a regression-free delivery. change_failure_rate remains unknown because complete post-delivery observation has not been established. Likewise, an individual intervention event does not establish complete observation of human assistance.

Event contract

The journal's existing correlation keys remain unchanged. Delivery identifiers live in event data:

{
  "type": "delivery.started",
  "context": {"task_id": "42", "run_id": "run-id"},
  "data": {
    "delivery_id": "episode-id",
    "contract_hash": "sha256",
    "ready_at": "2026-11-06T12:00:00Z",
    "ready_at_source": "observed",
    "interventions_observed": false
  }
}

Use ready_at_source: "dispatch" when historical Ready evidence is unavailable. This never counts as observed Ready coverage. delivery.accepted carries merge_sha, reviewed_head, evaluation_status, contract_hash and report_id. Its matching evaluation.finished event uses stage integration and includes criterion statuses, head, contract hash and duration.

Optional delivery.intervention events carry intervention_id, kind (policy_approval, clarification, correction or rescue) and action (started or finished). Only an explicit interventions_observed: true on the episode or coverage_complete: true on an observation establishes the coverage needed to calculate an autonomy rate.

Aggregation and presentation

Percentiles are calculated from samples. Run-insight distributions now include values_ms, and the dashboard merges samples across projects before calculating p50/p90. It never averages project percentiles. Multiple legacy summaries without samples show an unknown combined percentile.

The existing operational cards are labelled Attempt duration and Period cost / merge to distinguish them from delivery cohorts. Changing the reporting period can include an episode's older attempt costs, so delivery cost and period spend intentionally have different scopes.

The task inspector shows the recorded commit and each criterion's result. It independently checks the current task contract and available worktree head. A mismatch shows Outdated evidence; unavailable current evidence shows Historical evidence with the reason. These labels do not alter approvals or allow a recorded result to authorize a changed commit. Rubric judgments are displayed as advisory evidence and do not grant merge approval.

Use Open benchmark report… in Metrics to inspect a saved benchmark report.json. The importer recalculates descriptive comparisons without invoking agents or executing report commands. It shows evaluated and inconclusive trials, pass-rate intervals, cost coverage, timing and paired gains/losses. Open Artifacts opens the directory containing the selected report. Importing a report does not independently rerun its checks or verify the provenance of its recorded observations.

Verification

DeliveryMetricsTests covers cross-window retries and cost, reopenings, incomplete and abandoned spend, requirement coverage, parallel execution, policy approval separation, missing evidence, combined percentiles and integration-head binding. EvaluationPresentationTests covers combined samples, legacy summaries, asynchronous report refresh and changed task requirements.