Function apply_quality_guards

uncommon_route/router/quality.py:264–378

(
    candidates: list[str],
    *,
    mode: RoutingMode,
    tier: Tier,
    lane: CapabilityLane,
    capabilities: dict[str, ModelCapabilities],
    continuity_floor: ServedQuality | None = None,
    step_risk: str = "normal",
    agent_pressure: float = 0.0,
    is_agentic: bool = False,
    has_tool_results: bool = False,
)

def apply_quality_guards(
    candidates: list[str],
    *,
    mode: RoutingMode,
    tier: Tier,
    lane: CapabilityLane,
    capabilities: dict[str, ModelCapabilities],
    continuity_floor: ServedQuality | None = None,
    step_risk: str = "normal",
    agent_pressure: float = 0.0,
    is_agentic: bool = False,
    has_tool_results: bool = False,
) -> QualityGuardResult:
    quality_by_model = {
        model: model_served_quality(model, lane, capabilities.get(model))
        for model in candidates
    }
    target = target_served_quality(mode, tier)
    normalized_step_risk = str(step_risk or "normal").strip().lower()
    floor = minimum_served_quality(mode, tier)
    if lane is CapabilityLane.REASONING and tier is Tier.COMPLEX:
        lane_floor = ServedQuality.PREMIUM
    else:
        lane_floor = None
    if normalized_step_risk == "high":
        risk_floor = ServedQuality.BALANCED
    else:
        risk_floor = None
    hard_continuity_floor = continuity_floor if mode is RoutingMode.BEST else None
    effective_floor = stronger_quality(
        stronger_quality(stronger_quality(floor, lane_floor), risk_floor),
        hard_continuity_floor,
    ) or floor
    preferred_threshold = (
        stronger_quality(stronger_quality(target, effective_floor), hard_continuity_floor)
        or target
    )
    notes: list[str] = [
        f"lane={lane.value}",
        f"served-quality-target={target.value}",
        f"served-quality-floor={effective_floor.value}",
    ]
    if normalized_step_risk != "normal":
        notes.append(f"step-risk={normalized_step_risk}")
    if agent_pressure >= 0.35:
        notes.append(f"agent-pressure={agent_pressure:.2f}")
    if risk_floor is not None and normalized_step_risk == "high":
        notes.append(f"step-risk-floor={risk_floor.value}")
    elif risk_floor is not None:
        notes.append(f"agent-pressure-floor={risk_floor.value}")
    if lane_floor is not None:
        notes.append(f"lane-floor={lane_floor.value}")
    if continuity_floor is not None and hard_continuity_floor is None:
        notes.append(f"continuity-soft={continuity_floor.value}")
    exact_quality = stronger_quality(preferred_threshold, effective_floor) or target
    exact = [
        model for model in candidates
        if quality_by_model[model] is exact_quality