Issue: #752. Status: operator-threshold SLI evaluator and owner decision template; not GA.
The Auto Draft steward opens limited Draft PR shells; it never verifies code, marks a PR ready, or merges. Existing routine Actions artifacts are retained for a bounded 14-day window, which is not an archival guarantee.
| Indicator | Warning | Evidence | | — | — | — | | Scheduled job failure | recurring failures across scheduled runs | run IDs and timestamps | | GitHub API reserve block | multiple consecutive budget blocks | retained report counts | | Head-moved race | sustained increase above local baseline | recorded exact heads | | Already-existing PR race | sustained increase | bounded dedupe observations | | Evidence upload failure | any missing mandatory report | artifact inventory |
The owner must define numeric warning thresholds after observing a baseline. Unknown workflow data is UNKNOWN, not healthy. A CI queue backlog is a backpressure signal, not permission to bypass rate limits.
Before GA, explicitly record which governed owner decides (a) that 14 days is sufficient without long-term archive or (b) that selected evidence will be promoted to a durable store. Record classification, expected retention, privacy access, deletion policy and verification digests. Do not save every 15-minute report forever by default.
Move representative reports onto a different filesystem path/machine; validate with steward-report, rebuild steward-history, compare semantic/input-set SHA-256, render both local dashboards, and try a tampered report. Record actual commands, outputs and observed failure modes. This file does not claim that any drill ran.
GA requires observed branch-protection behavior, reviewed archive choice, portability drill and named SLI monitoring ownership; no simulated human approval, branch-rule bypass or assumed independent review can replace them.
idkmesh.steward_history.evaluate_steward_history_sli(history, ...)
accepts a strictly validated steward history and explicit owner-defined
thresholds for report volume, blocked run fraction, head-moved fraction, and
consecutive blocks. It binds its result to the history input-set SHA-256.
The status can be within_observed_thresholds, requires_attention,
or insufficient_evidence. Missing planned candidates leaves the
head-moved rate unknown rather than falsely healthy.
This diagnostic does not observe missing scheduled executions, failed artifact uploads, branch-protection behavior, retention ownership, or an executed disaster recovery drill. These stay explicit missing GA requirements. Class-based regression tests cover passing, failing, insufficient evidence, invalid thresholds, and tampered history. No integration authority is granted.