What Midgard scoresEach series' newest hour against its own hourly history.newestcommonnessshare ≥ this extrememodifiedZsize and sign
Ranking orderSTALE first, commonness ascending, then |z| descending.samples = 0commonness ↑|modifiedZ| ↓every row kepttry it ↓
Rows that misleadNotes flag thin history, MAD collapse, floor hits.n < 8 · score is noise|z| > 100 · flat, then movedcommonness ≤ 2/n · at floor3.5 ≤ |z| ≤ 100 · outlier
Dead collectorsCarried-forward ties look healthier; flagged STALE and sorted first.collector diesSTALEcommonness of newest hour
Read by valueInvariants and solvency always printed, whatever their rank.thornode_invariant_broken = 1thornode_solvency_excess_e8 < 0a permanently broken invariant scores as common
Whose metric is itNode-local, seen by one node, or network-wide.go_* process_*node-local*_consensus_*corroboratethornode_* pool_*network
health + tc-debugOptional step 1b sweep; tc-debug points here for vague symptoms.metrics.py --grep … --top 20--seriesLoki windowreport row
Go 1.26.10Toolchain bump across go.mod, CI images, Dockerfiles, scripts.1.26.91.26.10CI_TOOLS_VERSION 2 → 3 · image repin in a follow-up MR