Historical collection
The 09:00 UTC inspection workflow imports bounded Cloud resource history and available website-probe observations, then calculates forecasts. Missing history and incomplete imports are published as coverage notes. Availability uses minute bins across 30 days; only observed pass/fail minutes enter measured availability. Error budget is 43.2 minutes at a 99.9% target, while observation coverage remains a separate guardrail. Existing Cloud latency/freshness rules remain authoritative and linked; missing numeric targets are not invented.
Forecast validity
Resource histories are reduced to hourly bins for up to 14 days, requiring seven days, 80% coverage, and a recent sample. Disk growth must exceed 0.1 percentage point per day with R² at least 0.6 to produce a date. Sustained CPU/memory pressure means at least 20 of the last 24 hourly values at or above 85%. Queue depth is recorded where available; no unconfigured queue-pressure threshold is invented.
Recovery evidence
Workspace backup verification records encryption roundtrip and SQLite integrity outcomes from successful inspection workflows. Existing local and backend backup status, age, and scope metrics are displayed. A separate missing full-restore requirement remains until evidence is recorded. No retention policy for existing application backups is changed.
Fantasy Qwen production deadline
The target is established by fantasy_football_edge/importer/ai.ts at revision 7465011f375897592373589ed5451a563a956197: its production inference curl uses --max-time 220. The existing root log collector reads only GIN access records from fantasy-qwen.service, matching its current MainPID and exact POST /api/chat endpoint. Every observed completion in the rolling hour must be HTTP 2xx and less than 220 seconds; five completions and activity within fifteen minutes are prerequisites. No synthetic requests are generated. This measures inference HTTP completion, not answer quality or end-to-end fantasy-analysis correctness. The collector publishes only count, maximum duration, latest timestamp and outcome; process identity is used locally to invalidate evidence after restart. Observe reads these aggregates without journal privileges. Its existing performance check, weights, membership and freshness threshold remain unchanged.