Smart routing / 02

Route inside the rules.

Rispn reads the task, session, and economics, then chooses among administrator-approved candidates. When the evidence is weak, it keeps the requested model.

rispn.route / decision KEEP OR ROUTE
taskrepository_follow_up
requestedpremium_same_family
candidateapproved_efficient_model
policyALLOW
quality_evidencePOSITIVE
cache_stateWARM_CANDIDATE
decisionROUTE_CANDIDATE
Illustrative packetFull request remains unchanged
DefaultPass-through
ScopeOnly approved candidates
BiasAbstain on uncertainty
EvidenceEvery decision explained

The routing framework

Five gates stand between a request and a model switch.

A cheap model is not automatically a smart choice. Rispn routes only when policy, protocol compatibility, quality evidence, session state, and economics agree.

01PolicyIs this model allowed for this identity?
02CompatibilityCan it preserve the protocol and tool contract?
03QualityIs there positive evidence the task is safe?
04SessionWhat has happened in this coding session?
05EconomicsDo cache state and projected cost justify a switch?
KEEPRequested modelrisk, incompatibility, or weak economics
ROUTEApproved candidateall gates pass with positive evidence
ABSTAINRequested modeluncertainty is not a routing signal
One coding sessionState changes the decision
Turn 01Architecture planKeep premium
Turn 02Simple clarificationTry efficient
Turn 03Repository lookupStay efficient
Turn 04Failing testsRecover premium

Session-aware, not prompt-only

The same sentence can deserve a different model later.

Rispn considers recent errors, corrections, tool activity, warm-cache state, and prior route outcomes. Hysteresis prevents the router from bouncing between models without cause.

Quality before savings

The router needs positive safe evidence.

Absence of a warning is not enough. Rispn looks for an affirmative case that a compatible candidate can preserve the expected outcome. High-risk requests, active failures, and unclear tool contracts stay on the requested model.

SignalRouting effect
Short, direct follow-up in a stable sessionPositive evidence
Compatible model and response protocolPositive evidence
Active failing tests or recent correctionKeep requested
Unknown tool or response-format stateAbstain

Large-task decomposition

A large task should become reviewable before it becomes expensive.

For rare, high-confidence tasks, Rispn can pause with a visible plan. The user can continue directly or review how the work would be split. Nothing hides behind the conversation.

Large request Objective + constraints + expected validation Original request remains available
Rispn review gatePlan before spend
  1. 01
    Map the change surfacefiles, interfaces, risks, evidence
  2. 02
    Define bounded work packetsclear inputs, outputs, and dependencies
  3. 03
    Rejoin under one outcome gatetests and validation against the original goal
Continue directReview plan
Outcome contract One objective. One validation gate. Decomposition remains visible and opt-in

Explain every decision

A router should leave evidence, not mystery.

Rispn records the policy result, requested and selected model, route reason, cache state, provider-reported cost when available, and estimated alternatives with clear labels.

requested_modelpremium_same_family
selected_modelapproved_efficient_model
route_reasonpositive_safe_evidence
provider_selected_costOBSERVED
projected_alternativeESTIMATE
prompt_text_retainedNO
Pilot motion

Start in pass-through. Earn the right to route.

Observe your traffic, define the policy boundary, evaluate decisions in shadow, then canary live routing for an explicit cohort.

Request a routing pilot