Data Router

The Data Router answers one question: which source answers this question, under these constraints, at the cheapest sufficient tier [D] (data-plane.data-router, S4).

It is not the Capability Router. The Capability Router binds an executor to a task node. The Data Router binds a source to an information need. The corpus refuses the unqualified word Router precisely because these two decisions look alike and are not [A].

Admissibility runs before scoring

The ordering is the design.

1. Admissibility. Does the tenant's entitlement cover this source, for this purpose, in this jurisdiction, at this classification? An inadmissible source is removed here.
2. Scoring. Among the admissible, rank on authority, freshness, coverage, cost and latency.
3. Proposal. The router proposes the next tier. It does not authorise spending it.

Because admissibility runs first, an inadmissible source is never scored and can therefore never win on quality [A]. A system that scores first and filters afterwards will eventually show a user an answer it was not entitled to compute, and every individual line of that code will look defensible.

The stopping rule belongs to someone else

The Data Router proposes escalation to a more expensive rung. Only the Compute Governor authorises it, against one shared cost model [D] (brain.cost-engine, S4). One sufficiency question, one owner. Two homes for one calculation is how two calculations appear.

Cheapest sufficient, not cheapest

Cheapest sufficient is an immutable principle of the architecture, and the word that carries it is sufficient [A]. A router that optimises for cost alone on evidence-intensive work buys a cheap answer that fails review, which is the most expensive outcome available.

Acquisition cost sits inside the denominator of any efficiency measure, not outside it. On evidence-intensive work premium data can cost more than the inference performed over it, so an efficiency figure that counts tokens and ignores what the data cost is measuring the wrong quantity [A].