Causal State Machine

The Causal State Machine (CSM) is the connector between a Causaloid’s verdict and an effect on the outside world. It lives in deep_causality/src/types/csm_types and is built around two ideas: a state is “active” when its Causaloid evaluates to an active effect; an action is a function that runs when its paired state is active.

What it is

A CSM holds a thread-safe map of (CausalState, CausalAction) pairs:

pub struct CSM<I, O, C>
where
    I: Default + Clone,
    O: CsmEvaluable + Default + Debug + Clone,
    C: Clone,
{
    state_actions: Arc<RwLock<CSMMap<I, O, C>>>,
}

The constructor takes a slice of pairs. The map sits behind Arc<RwLock<...>>, so the CSM is shareable across threads and can grow or shrink at runtime through add_state, remove_state, and update_state, each implemented in its own submodule.

States and actions

A CausalState<I, O, C> is small. It carries:

• an integer id;
• a version for tracking iterations of the same logical state;
• a PropagatingEffect<I> of pre-bound data;
• the Causaloid that decides whether this state is active;
• an optional UncertainParameter for states whose Causaloid emits an uncertain effect.

A CausalAction is smaller still:

pub struct CausalAction {
    action: fn() -> Result<(), ActionError>,
    description: &'static str,
    version: usize,
}

action.fire() invokes the function pointer. The action surface is kept minimal on purpose: the conditional logic lives in the Causaloid, not in the action.

Evaluation

Two entry points are exposed:

• eval_single_state(id, data) runs the Causaloid for one state against caller-supplied data.
• eval_all_states() walks every registered state and runs its Causaloid against the data already bound to the state.

In both cases the effect returned by the Causaloid is inspected:

let is_active = match &effect.value {
    EffectValue::Value(val) => val
        .is_active(state.uncertain_parameter().as_ref())
        .map_err(CsmError::Causal)?,
    // Other variants (RelayTo, Error, etc.) are inactive for action firing.
    _ => false,
};

if is_active {
    self.fire_action_with_ethos_check(state, action, effect)?;
}

is_active is delegated to the CsmEvaluable trait. A deterministic effect resolves to true or false directly; an uncertain effect runs its hypothesis test against the state’s UncertainParameter. Anything that is not a value (relay, error, none) is treated as inactive and fires nothing.

When to reach for it

A CSM fits whenever the question “should this fire?” has to follow a causal verdict rather than a fixed threshold. Common shapes:

• sensor monitoring with thresholds that depend on context, not on the latest reading alone;
• alert routing where the alert depends on a pattern across signals;
• control loops in which an action should run only when an uncertain condition clears a confidence bar.

The CSM does not own a scheduler. The host application decides when to call eval_*. That keeps it cheap to embed in async runtimes, batch jobs, or hard real-time loops.

See also

• Example: sensor monitoring with CSM, backed by examples/csm_examples/csm_basic.
• Concept: Causaloid, the unit that supplies the verdict each state acts on.
• Concept: Effect Ethos, for policy checks that should gate an action before it fires.
• Concept: Uncertainty, for the hypothesis tests applied when a state’s Causaloid emits an uncertain effect.