Model Router

The failure it catches

Model routing is usually a chain of `if`s that grew one incident at a time, and its worst property is that a wrong pick is invisible: the reply still arrives, it still reads fine, and the only trace is on the bill or in a privacy review months later.

This is that decision as one pure function — classify the prompt, score the models you actually have, pick one, and say WHY, so the choice is auditable after the fact instead of inferred.

Modes

| mode | picks | |---|---| | `auto` | quality-first, light nudge from speed and cost, strong pull to local when the prompt looks private | | `quality` | strongest available for the task | | `price` | cheapest that still clears the quality floor | | `fast` | lowest latency that clears the floor | | `coding` | weighted on coding ability, not overall strength | | `research` | long context plus reasoning | | `privacy` | local only — a hard block, not a preference |

The three rules that matter

Privacy Mode refuses rather than degrades. No local model available returns `{ ok: false, error: "no_local_model" }`. It does not quietly use the cloud. A mode whose whole promise is "nothing leaves your machine" cannot have a fallback.

Sensitivity is inferred, and grounding counts. A prompt mentioning a diagnosis, an SSN, a bank account or a password is private — and so is any prompt grounded in your own documents, because that prompt now carries the corpus inside it.

A quality floor is what makes "cheapest" an answer. Without it, price mode is sorting by price and hands a contract summary to the weakest model on the list. When nothing clears the floor it still returns a decision rather than throwing out of the routing call.

Your catalog, your scores

The candidate list is passed on every call and per-model scores are overridable key by key. That is the only design that survives contact with reality: names, prices and tiers change every few weeks, and a router shipping a hardcoded table would be wrong within a month of purchase and wrong SILENTLY — still returning a confident pick for a model you no longer run. `DEFAULT_SCORES` is exported so you can read it, override it, or ignore it; an unknown model takes its tier's defaults rather than falling to zero.

What it does not do

It does not call any model — it returns a decision. It does not benchmark; the scores are a table you maintain. And it does not classify with an LLM, because a router that needs a model call to decide which model to call has a cost and latency problem in the one place you cannot afford one.

Verified

22 tests, 10 deliberate defects, all 10 caught — including Privacy Mode falling through to the cloud, the sensitive-prompt local boost quietly shrinking tenfold, price mode losing its quality floor, and caller score overrides being ignored in favour of the shipped table.

Delivery

Source delivered as a private repository invite within 24 hours of purchase. Single-product commercial license: use and modify in any number of products; no redistribution or resale of the source.

Interface

What you call, and what comes back. Types and signatures only — the implementation ships with the source.

  export function scoreFor( id: string, tier: 'local' | 'cloud', overrides?: Record<string, Score>, ): Score;
  export function classifyPrompt(text: string):;
  export function routeModel(opts: { prompt: string mode: RouteMode catalog: CatalogModel[] grounded?: boolean scores?: Record<string, Score> }): RouteDecision;