Host Coupling

Orrin treats the machine it runs on as part of its runtime context and learns that machine's normal behavior. The key idea: the same host metrics feed three deliberately separate mappings, so a small or busy machine is not mistaken for chronic distress.

flowchart LR
    HM[Host metrics: disk / swap / memory / battery] --> R[Reflex: absolute floors]
    HM --> S[Control signals: deviation from learned band]
    HM --> C[Cadence: absolute capacity]
    R --> RP[Pause heavy cycles]
    S --> SP[Bias internal state]
    C --> CP[Set cycle rate]

The three mappings

1. Reflex (absolute floors). The autonomic HostResourceGuard (supervisor/host_resources.py) watches disk/swap/memory below cognition and pauses heavy cycles at absolute safety floors — separate from the deliberative loop, because a thrashing loop can't be asked to rescue the substrate it runs on.
2. Control signals (deviation from a learned band). The resource self-monitoring layer feeds the same metrics into internal state on deviation — falling disk reads as constriction, rising swap as slowdown, a draining battery as a finite-horizon signal.
3. Cadence (absolute capacity). A small machine simply runs at a slower cadence, not a degraded one (resource_cadence.py).

Calibration and budget

brain/cognition/infancy.py learns that machine's normal oscillation before the runtime trusts its own deviation signals. A user-facing RAM budget slider (host_budget.py) feeds both the cadence policy and the reported "100%".

For the watchdogs and supervisor side, see Host Coupling & Supervisor.

Code pointers

• supervisor/host_resources.py — the reflex guard
• brain/cognition/host_resource_monitor.py, host_band.py, resource_cadence.py, infancy.py, host_budget.py