What the data shows

Published findings from Imantics research and deployments — every result labeled by what it does and does not establish.

The result we almost published — and why we didn't
August 28, 2026 — A result looked publishable until we asked what it actually established.

What a blind run-to-failure validation actually looks like
August 25, 2026 — What separates a blind run-to-failure validation from a demo: held-out data, no failure labels, a committed prediction, and a failure the model has never seen.

Running an LLM inside an air-gapped facility
August 21, 2026 — Cloud AI stops at the fab firewall.

Why zero-label matters for first-of-a-kind equipment
August 17, 2026 — First-of-a-kind equipment has no failure history anywhere.

The alarm-vs-warning distinction in water networks
August 12, 2026 — Threshold alarms fire when water is already high.

What autonomous should mean — the operator-controlled ladder
August 7, 2026 — Autonomous maintenance is not a switch, it is a ladder the operator climbs: recommend-only, approval-gated, autonomous — every step logged and reversible.

New findings by email

No pitch, just the data.