ROUTING-1 · A REGRESSION GATE, NOT A DISCOVERY
58 OF 60
CHOOSING-1 found that questions asking which form to use were being answered with a problem chapter, because both spaces use the same words. This asks sixty questions across twelve topics — the same vocabulary, five different requests each — and measures whether the kind of answer is identified before anything is retrieved.
Can the system tell which kind of answer is being asked for?
Almost always. On sixty questions built as twelve topics × five request kinds — choose a form, explain a form, explain a failure, ask about origin, ask about the system — the router identifies the kind in 58. Whether the reader then gets a useful answer is a different number: 40 of 60. The gap between those two is the finding, and it is not about routing at all.
READ THIS FIRST — WHAT THIS CANNOT SHOW
ROUTING-1 was not preregistered, and the person who wrote the router wrote the cases. It can show that a specific defect is closed on material the defect never saw. It cannot show that the router is good, and no claim of that kind is made here. CHOOSING-1 was the honest experiment; this is the gate that stops its finding from silently reopening.
EVIDENCE
Selection questions answered with a problem chapter
Zero of twelve. In CHOOSING-1 this was twenty-six of a hundred and twenty-four, and it was the reason for the change: the kind of answer is now decided before any record is matched, so a question about carrying content cannot be captured by a chapter about a failure merely for sharing the word "cite".
The router identifies the kind of request
58 of 60. The two misses are both a question about the system that names a form and reads like a request to explain that form — "what is a statement, in HAUSE?" — which is a genuine ambiguity in English rather than a routing bug.
Deciding the kind is enough to answer well
40 of 60. Routing improved and answering did not follow, because the spaces are not equally furnished: the kinds backed by records answer 24 of 24, and the kinds backed by hand-written keyword lists answer 16 of 36.
The spaces backed by records answer. The spaces backed by keyword lists refuse.
THE SAME QUESTION KINDS, RESOLVED TWO WAYS
ROUTED CORRECTLY ≡ ANSWERED USEFULLY
ROUTED CORRECTLY
classify() — what kind of answer is wanted
ANSWERED USEFULLY
the space that kind resolves in
24 / 24
24 / 24
PASS
form · history — resolved against the manifest and its origins
34 / 36
16 / 36
FAIL
select · problem · system — resolved against hand-written keyword lists
Routing is not the bottleneck any more. A question that reaches the right space and finds nothing there was never a routing failure — it is a space with a list in it where a record belongs.
ROUTED CORRECTLY ≡ ANSWERED USEFULLY — 1 of 2 rows agree; any disagreement is a distinct, named failure, never a warning.
USEFUL ANSWERS, BY REQUEST KIND
citation-select → choosing · citation-system → recommend-Provenance · provenance-problem → recommend-Provenance · provenance-system → no-form · certainty-select → no-form · certainty-system → form-statement · refusal-select → no-form · drift-select → no-form · drift-system → no-form · depth-select → no-form · depth-system → no-form · evidence-select → no-form · evidence-problem → no-form · evidence-system → form-evidence · comparison-select → no-form · legibility-select → no-form · legibility-system → no-form · terminal-problem → recommend-Terminal · history-select → no-form · history-problem → no-form
What replaces the keyword lists?
The grammar already holds thirty-five acts, each with the move it makes and the test that decides it — records, in the same shape as the manifest entries that answer twelve of twelve. Resolving selection against those instead of against a hand-written keyword list is the obvious next change, and this page deliberately does not make it: ROUTING-1 is frozen, and the change gets measured on fresh material rather than on the set that suggested it.
The experiment that found the defect, and the grammar the next fix would resolve against.
PUBLISHED 31 AUG 2026 · VERSION 1.0
CITEsite build 3ee5339 · built 2026-09-09
CITE THIS
Research note · 1.0
Hay, C. (2026). ROUTING-1 — deciding what kind of answer was asked for (Version 1.0). hause.design. https://hause.design/evals/routing-1
A gate, published like a result — including the part where routing stopped being the problem and coverage started.