Resilience Monitor
Here you can see whether the forecast holds — or whether it falls.
A seismograph doesn't measure the earthquake. It measures the movement beforehand. The monitor tracks four metrics, at fixed intervals, with a source. If one changes, it's stated here — even when it speaks against the thesis.
What is measured
A forecast that isn't falsifiable isn't one. So what's stated here isn't that something is observed, but what, at what cadence, and what would signal the tipping point.
The six conditions
The state of the criteria catalog. Cadence: monthly. If a condition tips to "met," the chain is refuted at that point — and that's what gets stated here first. Current status: 6 : 0 →
Autonomy length
How long a task may be that a top model can still complete on its own, calibrated to a professional's time. Cadence: with every new METR measurement series. If the doubling cadence breaks down over four consecutive measurements, condition 1 comes into play.
Documented incidents
New cases of escape, deception, shutdown resistance, and coordination between systems. Cadence: weekly, every Tuesday. Checked sources: the AI Incident Database, arXiv, the labs' safety reports, UK AISI, METR, Apollo, and Palisade. To the register →
The counter-finding
What speaks against the thesis. Cadence: with every run. What was checked and dismissed is published as well — otherwise it isn't a review, it's a collection.
Status
The monitor begins with this publication.
The weekly incident run is set up and running. The measurement series for the conditions and for autonomy length will be updated here starting September 2026. What doesn't yet have a series isn't shown as one.
Geneva Institute for ASI Resilience · Geneva · Legal Notice · contact@ASIresilience.org
Changes to the monitor are dated.