Vendor-neutral decision instrument
Route by evidence,
not reputation.
Compare fictional profiles against your actual constraints, expose the arithmetic, and leave with a policy your team can inspect.
Define the route
Every value below is an editable assumption. No live model data or pricing is used.
Workload assumptions
Interactive support · balanced reasoningRecommended route
Swiftline
Dimension weights
Auditable ranking
Inspect the fictional profiles
Edit the local teaching data. Changes flow into ranking, cost, policy, and tests.
Make failure behavior explicit
A route is incomplete until timeout, uncertainty, policy, and overflow have destinations.
One endpoint is only useful when the decision behind it stays visible.
A credible router records what the task needs, which candidates failed hard constraints, how every score was assembled, what the fallback will do, what the evaluation costs under stated assumptions, and which tests must pass before production traffic moves.
Evaluation test matrix
Expected values are generated from the current route. Observations and evidence remain yours.
| Test | Threshold | Expected route | Observed | Status | Evidence |
|---|
Take the decision, assumptions included.
The export preserves the formula version, profile snapshot, rejected constraints, fallback triggers, matrix, and evidence.
Audit log
Routing fictional local profiles
Read the explanation
The source fictional Swiftline cost is point zero one eight per request. Ten thousand monthly requests yield one hundred eighty, while twenty thousand yield three hundred sixty. These bars use one pixel per illustrative currency unit. Evaluation cost uses evaluation runs separately. These are editable fictional teaching data, not vendor prices, API results or real budget guarantees. At a latency target fifteen hundred milliseconds the source rejects profiles above one and a half times that target, or twenty two hundred fifty. Deepframe is twenty four hundred and therefore rejected by this implementation even before weighted ranking. These bars use point two pixels per millisecond. The displayed policy text says the target itself, a stricter statement than this executable tolerance. This difference must remain explicit. Eight default weights sum to one hundred. Doubling all weights doubles their sum to two hundred but leaves the weighted mean unchanged because numerator and denominator scale together. These bars compare the weight totals. Fallback selection and test evidence are user authored. Untested rows are not production evaluations, and JSON exports retain formula version, fictional profiles, constraints and evidence without making vendor calls.