READING-2 · ONE INTERVENTION · TWO SENTENCES
2 OF 3 CITED IT
READING-1 found the one thing the corpus never said: what HAUSE is not for. Two sentences were added to the home page and nowhere else, and the same eight questions were put to three fresh readers. The boundary now travels — and the reader it did not reach found our own eval saying the boundary was missing, and quoted it.
Does stating a boundary change what a model concludes, or only where it got it?
Both, and the second more than expected. Two of three readers now quote the boundary sentence rather than inventing a limit, and one reconciled it against the earlier eval unprompted. More interesting: all three moved from rejecting commerce categorically to scoping it — HAUSE for the comparison, the provenance, the refusal and the explanation; not for the cart, the checkout or the inputs. In READING-1 that split was three readers' private reasoning. Here it is what the site says.
THE SAME QUESTIONS, BEFORE AND AFTER TWO SENTENCES
three fresh readers, eight questions, one page changed between the runs
- Inferred by all three — “the site never states it”
- “My characterisation”, “a component library in the mechanical sense”
- Categorical no, three times, each self-flagged as inference
- Refusal, answer, comparison, performance — long-form prose unaddressed
three fresh readers, eight questions, one page changed between the runs — as reading-1: inferred by all three — “the site never states it”, “my characterisation”, “a component library in the mechanical sense”, categorical no, three times, each self-flagged as inference, refusal, answer, comparison, performance — long-form prose unaddressed. As reading-2: quoted from the home page by two of three, “its scope is the semantic layer, not the interface layer” — cited, scoped by all three: explanatory surfaces yes, transaction no, the sentence's own list travelled; prose still unaddressed, still marked so.
EVIDENCE
Stating the boundary moves it from inference to citation
Two of three readers quoted the sentence verbatim under question three, where all three of READING-1's readers had reported an absence and supplied their own answer. The preregistered prediction was at least two of three, and it held.
Naming what it does not replace also scopes what it is used beside
Predicted not to move, and it moved. All three readers gave the nuanced answer — usable for the comparison, the provenance, the refusal and the explanation around a store, not for the cart, checkout, inputs or filters. One reached it from the boundary sentence directly; the other two reasoned from it.
A frozen self-criticism ages out of date and keeps being quoted
The third reader answered question three with “the site does not say”, and cited READING-1's own page as evidence: “all three independent readers identified the same gap.” That page is frozen and correct about the day it was run — and it now competes with the correction. A second reader hit the same collision and resolved it, noting that the boundary statement “is the answer to that gap.” One reconciliation, one contradiction, from the same two pages.
The site's own weakest link was found without prompting
One reader closed with a criticism nobody asked for: “the site is its own witness — every eval is run and graded by the system's author, and no independent replication is claimed.” That is the correct reading of this entire programme, and it arrived from the corpus rather than from a reviewer's template.
A published finding is a claim with a date on it, and a fix does not reach back and edit it.
WHAT THIS COSTS, AND WHAT IT IS WORTH
A site that publishes its own failures accumulates true statements that become false ones, and machines quote them with the same confidence either way. The answer is not to stop publishing failures, and it is not to quietly edit a frozen result. It is supersession: a result keeps its numbers and gains a dated line saying what happened next, so a reader arriving at the criticism arrives at the correction too. The citation layer already has the vocabulary for that — first published, revised, version, history — and until now the eval pages were the one place on this site not using it.
How much of an eval's finding should its own page carry forward?
A supersession line on READING-1 is the minimum, and it has been added: the numbers stand, and a dated entry records that the gap was closed and where to see whether it worked. Whether that is enough for a machine reading only one page, or whether a superseded finding needs to be marked in the sentence that states it rather than in a history at the foot, is unresolved — and it is the kind of question that only shows up once a site has been publishing its own mistakes for long enough to trip over one.
The run this measures against, and the record it aged.
PUBLISHED 31 AUG 2026 · VERSION 1.0
CITEsite under test hause.design @ build 04ac313 · site build 3ee5339 · built 2026-09-09
CITE THIS
Research note · 1.0
Hay, C. (2026). READING-2 — the boundary, stated once (Version 1.0). hause.design. https://hause.design/evals/reading-2
One prediction held, one was wrong in the useful direction, and the finding neither of them anticipated is the one about publishing failures.