At each 30-min snapshot: r = workers_online / workers_active — score = r² (quadratic penalty).
workers_active excludes workers whose last snapshot is >7 days before the window end — stale machines don't count against the DC.
Continuity SLA = average of all snapshot scores × 100. Large outages are penalised heavily; small fractions are lenient:

ATO (Actual Transfer Out) = hashrate cmine commits to deliver to clients. If total HR ≥ ATO, clients receive 100%.
At each snapshot: snapshot_SLA = min(1, total_hr / ATO)
period_SLA = average(snapshot_SLAs) × 100
This combines availability and throughput — snapshots where HR = 0 score 0%.

DC share: avg_hr_dc / avg_total_hr computed only on online snapshots — avoids downtime deflating the share.
Throughput (when online): avg(min(1, hr_dc_ts / expected_at_ts)) — does the DC consistently deliver its proportional share?
Combined SLA: Continuity × Throughput — overall delivery including downtime.
Breach responsibility: during snapshots where total_hr < ATO (client impact), what fraction of the shortfall was caused by this DC?
dc_responsibility = max(0, ATO×dc_share − hr_dc) / (ATO − total_hr), averaged over all breach snapshots.
Lower = better. 0% = DC was not responsible. 30% = DC explains 30% of the client shortfall. 100% = DC was the sole cause.