Companion page to the interactive piece at othermode.ai/imagery. Wall v37, evidence page v5, 2026-09-03.
The Imagery Wall is an interactive piece about visual mental imagery: how vividly different people see a pictured object, and what happens when they are asked to change it. This page states what the piece claims, what those claims rest on, where the claims were wrong, and what has not been checked.
It is not a summary of the field. It is the working record of one non-specialist project holding itself to a stated standard, published so that the standard can be checked against the output.
A fuller per claim ledger exists in Chinese and is linked at the foot of this page. This page is a selection from it, not a replacement for it.
Every load bearing claim carries two independent marks, and from this version a third.
The third mark is new and is being applied to the Chinese ledger now. It exists because a single tick cannot distinguish "I read the abstract" from "I found the sentence". Two entries on this project were already sitting on that ambiguity: one qualitative claim rested on an abstract, and one citation chain was verified at one link and unverified at the next.
These are listed by failure type rather than by topic, because the types recur and the topics do not. All five were found by the project, not reported by a reader. None of them had a reader, because until now this page did not exist.
The evidence page stated that the general population averages roughly 57 out of 80 on the VVIQ, citing a mid range control group mean of 56.95. On checking, that group had been deliberately recruited to fall between 55 and 60. The number is the mean of a selected band, not of a population. It was removed. What survives is narrower: the distribution is right skewed, and around nine tenths of people are classified as typical in vividness. How "typical" maps onto the six panels of the wall is this site's own mapping, and the literature does not divide it that way.
A later round found that the discarded figure sits inside the range a review gives for large samples. The removal was still correct. A number can be approximately right and still have been asserted on evidence that did not establish it, and the second fact is the one this page is about.
The site used a figure of roughly 250 milliseconds for how fast a mental image fades. The sentence was traced verbatim to the introduction of a 2019 paper on children, which attributed it to a 1994 book without a page number. The book's own passage on maintenance is qualitative and gives no figure. The 250 millisecond value belongs to iconic memory and the interval between eye movements, which is the persistence of something just seen, not of an image generated voluntarily. The claim is now overturned and on the banned list. What survives is the mechanism without the number: an image decays unless it is actively refreshed.
That link has now been traced, and the answer is a split one. The 2019 paper's whole description of imagery stages is explicitly taken from a 2013 clinical-psychology review, and the sentence sits inside that borrowed passage. But the review's own account of maintenance is qualitative and carries no figure. The framework was inherited; the number was not. Where the number entered is still unknown, and the earliest appearance found is the 2019 paper itself, which is not a reason to call it the origin.
The prevalence section was suspended entirely on the grounds that the study it rested on was not a prevalence study but a survey of a self selected extreme sample. That was right about the study and too broad about the consequence. Two later rounds each found the previous round's reason too broad. The figures do have a traceable source: they are band proportions from the same sample, restated in the field's own review literature. The suspension of the section was right for a reason none of the three rounds stated, which is that the same numeral, 2.6%, denotes the low band in one reading of that source and the top band in another. The site now gives a range from several sources rather than any single number, because the figures diverge mainly by where the cut point is placed and by whether the sample was self selected.
Across several revisions, edits to the evidence page were applied by a script that matched a target string and reported success whether or not the match occurred. Three sections, including the page's own revision history, were absent from the file for a period during which they were described as present. No content was lost, because it survived elsewhere, but the completeness of that page was false while it was being asserted. Every edit is now checked for whether it actually landed, and the incident is recorded on the page it happened to.
An earlier version distinguished the two most vivid panels by involuntariness, and had built an interaction in which the image began to move on its own. That conflates voluntary vividness with intrusive imagery, which is studied in clinical populations under a different literature. Involuntariness has not been established as a feature of high vividness in non clinical people. The interaction was removed, and the wall now names this as a common claim it does not adopt.
The report uses one frozen rule. It compares only the direct answers stored as
q_change and q_change2, the same question asked before and after
the wall. If either raw field is missing, the result is missing. If either answer is
“Hard to say”, the result is ambiguous. Otherwise equal values are reported as the same
and unequal values as different. The endpoint control probe_end is shown only
as neutral context and does not classify the result. “No picture, I just knew” is a
valid direct answer with the same visual and analytic weight as every other answer.
This describes this run only. It is not a personality type, diagnosis, ability score, or validated measurement. The exact two answers and the research question are revealed only after completion.
Group context is shown only for an aggregate made for site v37, schema
15, the compatible frozen instrument and population
research_primary. A cell must contain at least 20 completed runs;
below that threshold it is suppressed and not displayed. Missing, slow, malformed, or
incompatible aggregate data also stays hidden. These are anonymous aggregate counts
from a self-selected sample, not a population estimate.
An older report compared endpoint controls and also mapped “done” incorrectly. That
display logic was a bug; it did not rewrite the stored answers. Compatible legacy site
v36 / schema 14 rows that retain the raw q_change and
q_change2 fields can be reanalysed with the frozen rule above. If either raw
field is absent, the old personal result cannot be reconstructed. Original rows are
never edited to make the correction appear retroactive.
Site v37 / schema 15 stores strict answer rows separately from optional comments and optional notification addresses. Funnel and share events are anonymous aggregate counters. A comment, an address, a share event and an answer are not joined into a participant profile. The requesting address is not stored in an answer row. Ephemeral anti-abuse buckets expire after 72 hours and idempotency keys after 48 hours. No deletion schedule has yet been set for anonymous research answer rows.
There has been no ethics review and no preregistered analysis plan. The two probe-form wordings have not been validated as equivalent across languages, and versions whose instrument or sampling changed must remain separable.
From v36e the first screen carries one sentence: "What people imagine may not be the same thing at all." This section says where it stands.
The claim: individual differences in imagery are large. ✅ Same verified set as the distribution sections: Wright et al. 2024, two samples (N=3,049 and N=9,063), roughly 0.9–1.2% aphantasia, ~3% hypophantasia, ~6% hyperphantasia, with about nine in ten in the middle band; Dance, Ipser & Simner 2022 (N=1,004), ~4% at the ≤32 cutoff; Zeman 2024 review, ~1% and ~3%.
The phrasing "not the same thing at all" 🔵 is this site's wording: it sets the two ends of the spectrum side by side; the qualitative and behavioural evidence for each end is in the sections above. "May" keeps it a possibility, not a claim about everyone.
Noted: the sentence appears before any imagining task, deliberately contains no apple, none of the three arrival-mode words, and sorts nobody into a type; it is a uniform exposure for the whole batch, recorded in the methods addendum.
Author. Li-Ning Wang, independent researcher, Taiwan. No institutional affiliation.
Funding. None. No grant, sponsorship, commission or institutional support. Self funded.
Competing interests. The author is the founder of Niaur Ltd, a company working on AI tooling in the governance and interpretability space. None of that tooling is publicly released and the company has no revenue generating product. This site carries no advertising, no sponsorship and no affiliate links, and its content is not reviewed or approved by the company or anyone acting for it.
Review status. This work has not been reviewed by an independent human expert. It has not been peer reviewed and no imagery researcher has read it. What happened instead is rounds of cross model checking against the primary literature, which overturned several claims previously treated as settled, including three of the five above. Cross model checking is not peer review, audit or validation. It catches errors and does not substitute for expert judgement, and the models involved share failure modes with each other and with the author.
AI collaboration. This site was built in collaboration with AI systems throughout, and that collaboration is itself an object of study for the author. The models wrote code, drafted copy, read sources and checked each other. The author set the direction, ruled every open question and is responsible for everything published. No AI system is credited as an author. Known limits, stated because they shaped the result: the models produced confident claims that later proved wrong, performed differently when asked to build than when asked to check, and attributed statements to the wrong speaker more than once. This page exists partly to make that visible rather than invisible.
Not claimed. The wall is not a diagnostic instrument and not a validated measure. The questionnaire it borrows its anchors from is not a clinical screening tool. Low vividness is an individual difference on a continuum, not a disorder. No claim is made anywhere that using this piece improves understanding or learning.
If you find an error here, reporting it is not a favour. It is the condition under which a page like this earns the right to be read. Write to lining.wang@niaur.com. Mail only, there is no form. Corrections are published alongside what they correct, credited to whoever found them if they wish.
This page has already corrected itself once. Version 1 stated that one of the disputed prevalence figures had no source. It has one, and always did. The sentence was written by restating a conclusion from the underlying ledger without rechecking it, which is the exact failure the read depth mark exists to prevent. Version 1 was never uploaded, so no reader saw it, and it is recorded here rather than quietly replaced.
Text and this evidence record are licensed CC BY 4.0. Code is licensed MIT. This is a project rather than a publication: there is no DOI and no frozen version, so please cite it as a website with an access date.