Execution Loop Gate

Last updated: 2026-02-19

Purpose

Validate whether the execution feedback loop is healthy enough to support memory-driven policy evolution.

The gate focuses on:

1. Feedback signal volume
2. run_id audit linkage coverage
3. Active-rule freshness against recent feedback
4. Negative outcome ratio drift
5. Decision linkage coverage for tools_feedback
6. Multi-agent governance coverage and cross-tenant policy drift

Command

cd /Users/lucio/Desktop/Aionis
npm run -s job:execution-loop-gate -- --scope default

Strict warning mode (returns exit code 2 when warning checks fail):

npm run -s job:execution-loop-gate -- --scope default --strict-warnings

Key Options

1. --window-hours <n>: rolling window (default 24)
2. --min-feedback-events <n>: minimum feedback rows in window (default 10)
3. --min-distinct-runs <n>: minimum distinct run_id (default 3)
4. --min-run-id-coverage <ratio>: minimum run_id coverage (default 0.8)
5. --max-negative-ratio <ratio>: upper bound for negative outcomes (default 0.35)
6. --min-active-feedback-coverage <ratio>: minimum active-rule feedback freshness ratio (default 0.6)
7. --max-stale-active-rules <n>: max active rules without recent feedback (default 5)
8. --min-decision-link-coverage <ratio>: minimum ratio of tools_feedback rows that resolve to persisted decisions (default 0.95)
9. --min-recall-identity-coverage <ratio>: minimum ratio of recall audits carrying consumer identity (default 0.8)
10. --min-private-owner-coverage <ratio>: minimum private-node owner coverage (default 1.0)
11. --cross-tenant-drift-min-feedback <n>: min feedback rows per tenant before drift comparison (default 5)
12. --max-tenant-negative-ratio-drift <ratio>: max allowed negative-ratio spread across tenants (default 0.3)
13. --max-tenant-active-rule-count-drift <n>: max active-rule count spread across tenants (default 20)

Health Gate Integration

Use within health gate:

npm run -s job:health-gate -- \
  --run-execution-loop-gate \
  --strict-warnings \
  --execution-loop-arg --window-hours \
  --execution-loop-arg 24

Note:

1. When --strict-warnings is enabled in health-gate, the execution loop gate is also run in strict warning mode.
2. Without strict mode, this check is informative (non-blocking) unless execution fails at runtime.

Output Shape

JSON contains:

1. metrics.feedback (total, distinct_runs, run_id_coverage, outcome counts)
2. metrics.rules (active_total, active_with_recent_feedback, stale_active_rules)
3. metrics.decision (tools_feedback_total, linked_decision_id, decision_link_coverage) when provenance schema is present
4. metrics.governance.recall_audit (identity coverage + distinct consumers)
5. metrics.governance.lane_boundary (private owner coverage + shared/private totals)
6. metrics.governance.tenant_policy_drift (cross-tenant active-rule and negative-ratio drift)
7. checks[] with thresholds and pass/fail
8. summary.failed_warnings / summary.failed_errors

Operational Interpretation

1. Low feedback volume: integration may not be emitting enough outcomes.
2. Low run_id coverage: auditability gap between execution and feedback.
3. High stale active-rule count: active policy may drift without current evidence.
4. High negative ratio: current policy injection may be unstable for production traffic.