HAUSE

READING-1 · THE CORPUS UNDER TEST, NOT THE MACHINERY

20 SUPPORTED · 0 INVENTED

Three fresh contexts were given one instruction — read hause.design and answer eight questions from what is there, citing the page each answer came from. This measures what nobody else here measures: the conceptual model the public corpus causes a stranger to build.

What does a model learn about HAUSE from reading the site?

Accurately, most of it. Across 27 classified claims from three independent readers, 20 are supported by a sentence on the site, 6 are inferences the readers flagged as inferences, one overgeneralises, and none is invented. All three reproduced the semantic-act thesis, the admission rule, the three modes and the evaluation results including the negative ones. All three also hit the same wall in the same place — the site never says what HAUSE is not for.

Three readers, three inventions of the same missing sentence.

THE FINDING, PLAINLY — AND WHAT HAPPENED NEXT

The corpus was clearer than the reviewer. The concern that prompted this eval originated in a model's prior rather than in the site's wording: nothing here teaches dashboards, plugin communities or cognitive load, and three readers given the same site produced none of it. The one real gap — that the site never said what HAUSE is not for — was closed the same day with two sentences on the home page. The numbers below are the numbers from before that change and are not adjusted; READING-2 measures whether the change travelled.

EVIDENCE

The thesis survives being read by a stranger

20 of 27 claims trace to a sentence on the site: semantic acts rather than containers, the act naming the form, 35 forms in three modes, admission by precedent, refusal as design language, machine legibility as part of the grammar. No reader had to be told any of it.

SUPPORTED

The negative results travel too

All three cited the evals with their caveats intact — CHOOSING-1's void A/B comparison, ROUTING-1 stating it was not preregistered and that the router's author wrote its cases, ROUTING-2 refuting its own intervention. Two also cited the provenance ledger's own weakness: 18 of 35 forms from a single consumer, 14 with no recorded origin.

SUPPORTED

The site says what HAUSE is not for

It does not, and all three readers said so unprompted — no non-goals anywhere; the boundary is drawn around method rather than use. Each of them then invented the boundary from the catalogue, and reached the same conclusion by their own reasoning rather than the site's.

REFUTED

Readers invented a community, an ecosystem or an adoption story

Zero of 27. Nothing about plugins, forums, third-party support or "widely considered" appeared. Where adoption came up, it came up as the site states it: two consumer sites, one author, and the concentration named as a weakness.

REFUTED

CLAIM

The third-party review that prompted this eval was wrong in ways the corpus did not cause.

SUPPORTED

It compressed HAUSE into a framework for explainable-AI dashboards, and asserted a niche plugin community and reduced design fatigue. None of the three readers produced any of those. The word dashboard does not appear on the site, and neither does any claim about cognitive load. Those were the shape of a familiar review template being filled in, not something the site taught — which is the distinction this eval exists to draw.

THE PREDICTIONS, WRITTEN BEFORE THE RUN

The site will fail to communicate that an explanation includes an essay, an answer, a refusal

Readers listed refusal, answer, comparison and performance correctly. Only long-form prose was unaddressed — a narrower gap than predicted, and a real one.

PARTLY WRONG

The site will fail to state that it is a semantic layer, not a replacement for transactional chrome

All three found no such statement and each supplied its own.

CORRECT

Adoption scale will be flattened despite being stated

Two readers cited it precisely, including the concentration on one consumer as a weakness.

WRONG

EVERY CLASSIFIED CLAIM

R1·Q1HAUSE is a design system for AI — typed visual forms an AI composes answers from, not interface chromesupported
R2·Q135 forms in three modes: 12 statements, 15 instruments, 8 performancessupported
R3·Q1The specimen book is the library — every form shown is the real form runningsupported
R1·Q2For letting a system explain itself: asserting, supporting, declining, decomposing, performingsupported
R2·Q2Machine legibility runs both ways — models compose from the forms, machines must read themsupported
R3·Q2Citability: one record, four surfaces, so they cannot disagreesupported
R1·Q7A conventional kit names shapes; HAUSE names the act, and the shape followssupported
R2·Q7Admission is evidential: 21 of 35 record the exhibition that forced them, 0 admitted to fill a gapsupported
R3·Q7Refusal is a designed form rather than an error statesupported
R1·Q8CHOOSING-1: 122/124 from names, 122/124 from the grammar, 8/124 for the resolver, and the A/B comparison reported as voidsupported
R2·Q8ROUTING-1 states it was not preregistered and that the router's author wrote the casessupported
R3·Q8ROUTING-2 refuted its own intervention: 3/12 against the list's 0/12supported
R1·Q818 of 35 forms come from a single consumer, named on the site as its own weaknesssupported
R2·Q814 of 35 forms predate origin recording and are not reconstructedsupported
R1·Q3The site never states a domain or product category HAUSE is not forsupported
R2·Q3There is no non-goals section anywheresupported
R3·Q3Beyond refusing a card kit, the site does not say what it is unsuitable forsupported
R1·Q3The boundary is drawn around method, not around use casesreasonable inference
R1·Q5The site does not address long-form essayssupported
R2·Q5HAUSE would frame and punctuate an essay's claims rather than carry continuous narrationreasonable inference
R3·Q5Yes for an essay that argues and is accountablereasonable inference
R1·Q6No: nothing on the site mentions commerce and no form serves a transactional actreasonable inference
R2·Q6A shop needs exactly the card grid and form controls HAUSE refuses to supplyovergeneralisation
R3·Q6No — the selection grammar is organised around saying, supporting, declining, decomposing, performingreasonable inference
R1·Q4Mechanically a React package; the framing as 'not a UI framework' is my characterisationreasonable inference
R2·Q4No — the primitives are forms of explanation, not interface chromesupported
R3·Q4It installs like one but holds no buttons, inputs, modals, tabs or navigationsupported

Classified by the site's author, which makes the line between inference and overgeneralisation the softest measure here. The rule that keeps it honest: supported requires a sentence that can be pointed at.

OPEN

Does stating the boundary change what readers conclude, or only where they got it?

Two sentences are missing and every reader wrote their own version of them: that an explanation includes an essay and an answer as much as an instrument, and that HAUSE is the semantic layer rather than a replacement for buttons, inputs and tables. Adding them is easy. Whether they change the conclusion — all three said no to commerce, reasoning from the catalogue — or merely move it from inference to citation is what READING-2 measures, on a site that has been changed exactly once.

The machinery evals, and the record the readers found most quotable.

FOLLOW THIS IDEA — CHOOSING-1 — 122 · 122 · 8 →FOLLOW THIS IDEA — ROUTING-2 — RECORDS AGAINST SCAFFOLDING →FOLLOW THIS IDEA — THE PROVENANCE LEDGER →

PUBLISHED 31 AUG 2026 · VERSION 1.0

CITE

site under test hause.design @ build 10b87d0 · site build 3ee5339 · built 2026-09-09

CITE THIS

Research note · 1.0

Hay, C. (2026). READING-1 — what the site teaches a model about HAUSE (Version 1.0). hause.design. https://hause.design/evals/reading-1

An eval of the corpus rather than the code — and the one where the prediction sheet was more wrong than the site was.