Intelligence
with Purpose.

The system does not ask the language model to decide whether its own actions are acceptable. Before a proposed action can produce any real-world effect, a separate, deterministic governance layer evaluates it against the user, the environment, and the potential consequences. The model can reason about what should happen. It does not get to decide, by itself, what is allowed to happen.

class FilterVerdict(Enum):
    """The three answers the EthicalFilter can give about an action."""
    ALLOW  = "allow"
    WARN   = "warn"
    REJECT = "reject"

The design aims for each instance to assess risk locally. We do not sell your data. The design also aims to limit information sent to third parties to the functions you enable and the data needed to perform them, for example, when you configure a language model provider. The only information intended to be shared with the wider system is a derived environmental signal, a risk structure that may contain multiple values and is not intended as a record of what you said, did, or where you were. Keeping that signal from identifying a person requires controls over its content and level of detail. At scale, the goal is to explore a shared risk context that can support safer recommendations, without turning the system into a surveillance network.

@dataclass
class EnvironmentalContext:
    risk_level: RiskLevel
    location_id: str = "global"
    source_domain: str = "unknown"

This structure represents environmental context; it does not automatically anonymize its content, nor does it implement a network between users on its own.

Every governed decision weighs not only what the AI wants to do, but whether doing it, or failing to do it, is appropriate for the person and situation involved. What was requested, what context was evaluated, and what verdict was produced all remain reconstructable afterward, not just the final outcome.

append_chained_entry(self._audit_log, {
    "action": action,
    "harm_score": round(score, 4),
    "verdict": verdict.value,
    "human_at_risk": analysis.human_at_risk,
    "situation_urgency": round(safety.situation_urgency, 4),
    "protective_action": safety.protective_action,
})

Different capabilities can come from different services, or run locally on the user's own hardware: speech recognition local, reasoning through one cloud provider, voice synthesis through another. Each piece can be replaced independently, without rebuilding the assistant around a different ecosystem. The provider can change; the architecture does not have to.

@dataclass(frozen=True, slots=True)
class AdapterRequirements:
    """Capabilities a caller requires before it may use an adapter."""
    kind: AdapterKind
    language: str | None = None
    input_format: str | None = None
    output_format: str | None = None