living experiment 01 / visual interruption

The Change You Called in Advance

A 1997 paper printed 1.4 alternations and 0.9 seconds for its no-blank check; this page rechecks that arithmetic as 0.896 seconds on an inferred 0.640-second timebase, then runs a narrower paired visual task with an anonymous, four-integer, button-press-only intake.

Arm A: published-summary auditArm B: collecting, zero is visibleArm C: none, by design
The operable base

Predict first. Then look.

Two abstract scenes will alternate. One trial inserts grey blanks and one does not. One change exists in each. Your first click or tap stops the search, then you classify the change as presence or absence, or location.

Seal the answer before the scenes

When a grey blank separates the views, will you correctly locate and classify the change within 16 alternations, which is 10.24 seconds?

Your prediction

Nothing has been assigned or sent.

A reduced-motion preference is honored by keeping the flicker instrument closed. The rest of the page remains available. The task also needs a viewport of at least 720 by 480 to qualify for a contribution.

Arm A / the dead

One printed mean. One recoverable check.

Rensink, O’Regan, and Clark reported that their no-blank condition required a mean 1.4 alternations, or 0.9 seconds, on page 370. Their row-level observations were not found. This page therefore audits a published summary, not participant data.

Computed from transcribed cells

computing2 × (240 ms image + 80 ms blank) gives the 640 ms switch interval

Printed on page 370

computingpublished aggregate, not raw data

Checking the anchor.

The arithmetic multiplies the printed 1.4-alternation mean by an inferred 0.640-second timebase. The article prints 640 ms for the flicker cycle but never prints the no-blank frame schedule, so carrying that timebase over is an inference, supported by the printed rounding itself (0.9 divided by 1.4 is 0.64), not an independent reconstruction. The product, 0.896 seconds, rounds to the printed 0.9. The same paragraph reports p > 0.3 for its no-blank comparisons. That inferential clause cannot be regenerated from one mean, so it is transcribed as text and never presented as recomputed evidence.

The often-shared 7.3 versus 17.1 alternation contrast is not this page’s anchor. It depended on the unavailable complete normed 48-scene set and a separate five-describer central-versus-marginal classification.

Arm B / the living

Zero is a result state, not an empty box.

Checking the live store.

0store rows
0protocol-eligible rows
0out-of-protocol rows

No rows loaded.

What the bands test, exactly

Each primary paired score lies from -1 to 1 and is scaled into the shared betting confidence-sequence engine at alpha 0.025 each, so Bonferroni gives simultaneous 95% familywise coverage for the two primary sequences under their assumptions. A confidence sequence is valid at every moment simultaneously; a fixed-n 1.96 interval would quietly spend its error promise again whenever a visitor looked.

Each bankroll is an anytime-valid test of one sharp null. For co-primary 1 the null says the blanks shift no accepted contribution's detection probability; under that null, with the registered random swap of scenes between conditions, every scaled score has conditional mean exactly one half at every step, which is what the betting construction requires. For co-primary 2 the null says no accepted contribution's prediction probability differs from its blank detection probability. Each confidence sequence covers one common conditional mean, assumed identical across accepted rows; if that mean drifts with time or crowd composition, the sequence tracks no single number, can exclude the running crowd average, and can even go empty. That identical-conditional-mean assumption is stated here, not tested.

The three binary bands are separate descriptive 95% confidence sequences under the same fixed-population assumption, and they do not jointly have 95% coverage. Repeated looks are handled; self-selection, dependence, bots, a changing readership, unverifiable client-side assignment, and a manipulated sequence are not.

The check

What can fail here.

PUBLISHED ANCHORThe page loads timing and aggregate cells with printed page numbers, recomputes 640 ms and 0.896 s, and compares the transcription against both the arithmetic and the sealed analysis module's own frozen copy of the printed cells. One wrong cell in either document makes this panel red and closes the living display. This is a published-summary consistency check on an inferred no-blank timebase, not a reconstruction of observations.
ANALYSIS SEALAnalysis version checking must hash to 85f89eacbbcadf793ff5b87f7676d5c853319fdcac079e361a747032eaf52da1. The page checks the live bytes against this literal, rather than printing a fresh hash.
STIMULUS AND DATA SEALThe frozen stimuli, masks, protocol, and transcription must hash to e76ef051405292b63236bef4797286df70d2913d85e33b3c07f0bd4720e74784. This extra seal is necessary because an unsealed changed mask could alter outcomes without touching the analysis.
OFF-PAGE CHECKWaiting.loading command
loading command
EMPTY STATEThe arm opened on 2026-08-21. With no eligible rows it names zero, the opening date, and exactly what will appear. Under 12 rows it shows counts only. At 12, with each swap direction and each order holding at least two eligible rows across at least two scene sets, it opens the registered sequences and diagnostics; without that support it says so and shows counts only.
COLLECTION AND CONSENTThe intake is governed by the site's standing living-experiments privacy design: four bounded integers, no free text, no identity, day granularity, and nothing sent without this page's one deliberate button press. There is no external ethics body for this site, so no further sign-off exists to wait for; the versioned determination is recorded in the research folder. Named caveat: no independent side-by-side discriminability pilot preceded launch, so the per-scene table below is the accumulating substitute and every registered claim stays within-reader.
WIRE LIMITThe fixed API stores only prediction, two alternation counts, and one assignment cell. It cannot preserve timing traces, viewport failures, prior exposure, separate clicks, or duplicate identity, and it cannot prove a stored row used the honest client-side randomization. The analysis excludes only values outside the registered 16-alternation and 16-cell subranges, and does not invent a screened dataset.
FREE CHOICESThe cap is 16 alternations; the early threshold is 12; the support minimum is 2 rows per swap and order margin across at least 2 scene sets; paired primary alpha is 0.025 each; descriptive alpha is 0.05; the grid has 1,001 points; misses count as 16 only in capped-time summaries; four fixed scene sets are equally weighted in one sensitivity display.
WHAT IT CANNOT CONCLUDEA miss does not show that a visual representation was absent. It could decay or be replaced, lie in an inaccessible pathway, use a format the comparison cannot use, or never be compared. This task cannot establish eyewitness reliability, driving risk, attention sufficiency, or performance on unexpected real-world changes.
POPULATION LIMITAn anonymous, self-selected web sequence is not representative. Local storage is weak duplicate friction. The edge may rate-limit, but the research payload contains no IP or device field. Infrastructure logging is governed outside this instrument by the site privacy policy.
ARM A LIMITThe source gives an aggregate mean and prose about p > 0.3, but no observations that recover the response distribution or tests. Access to the PDF and demonstrations is not a data licence. This page redistributes neither the original scenes nor a guessed dataset.
NO MODEL ARMA text model has no visual stream. A vision model given both images tests image differencing; a video input tests frame sampling and system architecture. Neither is the same masked-attention protocol, so no machine crowd is manufactured.
Primary record and audit trail

The check leaves the page.

Rensink, O’Regan, and Clark, 1997, author-hosted article PDFPrinted pages 369-370 support the 240 ms image, 80 ms blank, repeated-state schedule, 48-scene method, 1.4-alternation no-blank mean, 0.9-second rounded time, and p > 0.3 clause. Read directly on 2026-08-21.
Rensink flicker demonstration pageChecked for demonstrations and resource links. It does not expose the participant observations needed to reconstruct the selected result.
Simons and Rensink, 2005Printed pages 18-19 support the four requirements of scope summarized in the check panel. The full text was read directly.
Howard, Ramdas, McAuliffe, and Sekhon, 2021Time-uniform confidence sequences are the methodological basis for remaining valid under continuous looks. The implementation itself comes from the site’s shared betting kit and its grid is disclosed.
Jagatap and colleagues, 2021An open later masked-flicker study with different stimuli, 250 ms timing, and a 60-second cap. It supports the broader paradigm but is not substituted for the unavailable 1997 observations.

Search limit: the scout checked the author site, article, title and DOI searches, repository-style searches, and Crossref metadata through 2026-08-21. The claim is only that no original row-level deposit was found in those places by that date.