Configuration reference¶

Calibration file¶

version: demo-baseline-2026-09-04
temperature: 1.0
clarify_threshold: 0.55
min_retrieval_score: 0.08
tiers:
  low:
    route: 0.55
    margin: 0.08
  medium:
    route: 0.65
    margin: 0.12
  high:
    route: 0.76
    margin: 0.18
  critical:
    route: 0.90
    margin: 0.25

These are demonstrative baseline values, not production recommendations.

Field	Type	Constraint	Meaning
`version`	string	Required	Immutable calibration bundle identity
`temperature`	float	Greater than 0	Temperature scaling applied in logit space
`clarify_threshold`	float	0–1	Minimum confidence for a `CLARIFY` decision
`min_retrieval_score`	float	0–1	Minimum first-stage candidate score
`tiers.*.route`	float	0–1	Minimum route confidence for that risk tier
`tiers.*.margin`	float	0–1	Minimum gap over the second candidate

All four risk tiers must be present. Route thresholds must not decrease from low through critical.

Routing policy¶

from flowroute import RiskTier, RoutingPolicy

policy = RoutingPolicy(
    allowed_risk_tiers={
        RiskTier.LOW,
        RiskTier.MEDIUM,
        RiskTier.HIGH,
        RiskTier.CRITICAL,
    },
    reject_reasoning_requests=True,
    reject_multi_action_requests=True,
)

Field	Default	Meaning
`allowed_risk_tiers`	All tiers	Global risk eligibility
`reject_reasoning_requests`	`True`	Abstain on common analysis, advice, and planning language
`reject_multi_action_requests`	`True`	Abstain when multiple actions are detected

Router¶

router = FlowRouter(
    registry,
    retriever=my_retriever,
    verifier=my_verifier,
    calibration=my_calibration,
    policy=my_policy,
    top_k=8,
)

top_k must be between 1 and 64.

Runtime modes¶

from flowroute import RouterRuntimeConfig

runtime = RouterRuntimeConfig.production(
    max_context_bytes=64 * 1024,
    allow_exact_event_routes=False,
)

Field	Production value	Meaning
`require_catalog_version`	`True`	Reject requests without active catalog lineage
`require_workflow_allowlist`	`True`	Require an authorization-derived candidate set
`allow_exact_event_routes`	`False` by default	Do not trust caller-supplied event types as routing authority
`allow_debug`	`False`	Do not expose candidate-level scores
`fail_closed_on_backend_error`	`True`	Convert model errors to a typed abstention
`max_context_bytes`	`65536`	Bound the serialized context size

Set allow_exact_event_routes=True only for an authenticated structured-event entrypoint. The setting is bound into the production artifact manifest.

HTTP settings¶

from flowroute.api import ApiSettings

settings = ApiSettings.production_defaults(
    trusted_hosts=["router.example.internal"],
    cors_origins=["https://console.example.com"],
    max_body_bytes=128 * 1024,
)

Production settings require an authorization callback, disable interactive API docs, reject wildcard hosts and origins, and validate host/origin syntax. TLS, authentication, global request deadlines, and distributed rate limiting remain at the ingress or service-mesh boundary.

Catalog limits¶

The typed schema supports:

up to 50,000 workflows per snapshot;
up to 30 positive capabilities, exclusions, inputs, preconditions, or event types per workflow;
up to 50 examples per workflow; and
contract descriptions up to 2,000 characters.

Schema capacity is not the same as tested performance. The lexical baseline prepares its TF-IDF index once, but it is still not appropriate for a large production catalog. Precompute learned embeddings and use indexed retrieval for scale.