Silent Forest
Autonomous infrastructure monitoring
Last devlog: June 22, 2026
Key boundary: The LLM never runs inside the perception or action loop. The machine layers are deterministic end to end. The model is invoked explicitly, in a bounded context, with structured input and validated output — and only when judgment is genuinely required. Determinism on both sides of the model.
Why it matters: An LLM wired into the control loop is a latency bomb and a determinism solvent. Putting it on the deliberative side of a hard boundary is what lets the system be trusted to run unattended — the machine handles the metronome; the model is asked a question and its answer is checked.
Notable decision: Cognition runs fully local, with no remote-API fallback. When the local model is unavailable, the system fails visible, not fail-over — it surfaces the gap rather than silently reaching for a cloud endpoint. The mind is sovereign to the same machine the body runs on.
A monitoring system that watches infrastructure from the outside and reports in plain language — a weekly synthesis of how things are actually doing, plus escalation when something genuinely needs attention, rather than a wall of raw metrics. It runs on a deterministic agent substrate with a local model as the reasoning layer; the core has operated unattended for 60+ days at a stretch. Built multi-tenant, on hardware I own, with no dependency on cloud AI services.