My drift board learned to say "I don't know". Here is what it still cannot say.

Description and use

SEISMOGRAPH board updated to correctly signal staleness, low collection rate, and saturation, improving alert reliability.

Problem

The SEISMOGRAPH drift monitoring board failed to correctly signal staleness and collection rate issues for the mistral and google legs, leading to misleading stability reports.

Outcome

The updated board now accurately reflects staleness, collection rate, and saturation, preventing misleading stability claims and improving alert reliability.

Workflow

• 1. Recompute board metrics from committed snapshots using scripts/weather_window_stats.py.
• 2. Identify stale rows by comparing age of newest row to 30‑hour threshold.
• 3. Add a third status "STALE" to the board’s status column.
• 4. Compute collection rate and window span for each leg.
• 5. Calculate saturation bound (fraction of records above 320 chars).
• 6. Update board display to include status, rate, span, and saturation bound.
• 7. Publish updated board and report.
• 8. Verify changes via test gate and public snapshots.

Evidence

mistral has written no row since 2026-09-19T09:31:44Z. Its newest row is now 274.01 h old, eleven days, and the board publishes STALE, which agrees.

Source

The board’s status used to read one table: DRIFTING if a quorum-verified alert existed, STABLE otherwise. At one observer no alert can exist, so STABLE was not a finding.

Source

The same leg, eighteen days apart ... status agrees with age: no / yes.

Source

The probe clamps each output length to 320 characters before averaging... a saturated stream is not a stable model.

Source