Routing

Data pipeline and router shadow

How Conflux collects safe learning signals before enforcing auto routing.

Purpose

Data Pipeline v0 and Auto Router v0 collect explainable request metadata before Conflux changes any live route. The first release classifies requests in shadow mode and stores metadata-only dataset examples for later compressor and router evaluation.

Shadow mode

No route changeThe existing resolver still selects the actual model and provider route.

Task categoryRequests are classified into work types such as debugging, planning, SQL analytics, incident ops, or security review.

Decision metadataEach shadow event stores difficulty, risk, long-context need, current-info need, recommended tier, confidence, and reason codes.

Audit signalactualRouteUnchanged stays true so analytics can prove the classifier only observed the request.

Dataset rows

PromptDatasetExample rows are metadata-only by default. They store hashes, counts, feature JSON, label JSON, eligibility status, and redaction metadata.

ROUTER_SHADOW
source: auto-router-shadow
label: category, tier, confidence, reason codes

COMPRESSOR_REDUCTION
source: prompt-optimizer
label: changed, saved chars, reducer ids

Safety contract

The v0 pipeline does not store raw prompts, latest user prompts, provider payloads, or optimized prompt bodies as training examples. Raw-text teacher labeling requires explicit workspace opt-in, redaction, retention rules, and review before it can feed a model.