Not current Houston, not metropolitan Houston, and not a later estimate. The date, incorporated place, and decennial count make the target fixed.
First, one estimate
Before the evidence.
Your local participation route is already spent.
This browser says it contributed or chose to view. That marker is only advisory and can be cleared. The server separately limits accepted rows.
You may contribute one innocuous numerical estimate. Do not search. The store receives only four bounded integers: the random condition, your estimate, whole seconds elapsed (capped at 600), and your exposure answer. It receives no free text, demographic field, location, fingerprint, or raw user agent from this form.
Start makes a cryptographic random low or high assignment in this browser, then hashes its condition and nonce before showing the prompt. The collective API has no signed-token field, so this is a local commitment, not a server-signed proof or a guarantee of one human per row.
Assignment commitment, made before this prompt
pending
Allowed range: 0 through 8,000,000.
Exactly what will be sent
Nothing has been sent yet. Pressing the next button is the deliberate send.
Your local assignment receipt
Condition:
Exact preimage
SHA-256
Hash the UTF-8 preimage and compare it with the digest that appeared above the prompt. This proves the browser fixed this condition before display if you trust the page code that produced it. It does not prove server assignment, identity, or honest execution on a modified client.
The fixed target
One count, with its borders left on.
No diagnosis is hiding here
Your other possible response was never observed. The experiment can estimate an average effect of the whole low-versus-high prompt package. It cannot tell whether you personally were anchored, nor identify adjustment, priming, irrationality, or any other mechanism.
Lineage, not substitution
The famous result has only two surviving rows.
1974 SUMMARY ONLY
The Science article reports two medians for its wheel demonstration. It does not report the sample size, group sizes, participant estimates, variance, histogram, or an inferential test for that demonstration. No participant distribution is reconstructed here.
These are medians, not means. The primary article says numbers were determined by a wheel and different groups received different numbers. It does not document the wheel’s construction or call it rigged. The accuracy-payoff sentence has no numeric comparison beside it.
ARM A: 2014 REPLICATION DATA
Chicago, recomputed from 5,284 licensed observations
This is Many Labs 1, not the 1974 original. Its cleaned Chicago rows survive under CC0. Exact duplicate condition, estimate, and deposited-rank triples are losslessly collapsed into frequency cells in page-data.json.
| Quantity | Low prompt | High prompt | Aggregate |
|---|---|---|---|
| Recomputed medians and n | loading | loading | loading |
| Printed record | median not printed there | median not printed there | loading |
Checking the shipped cells.
The historical calculation uses the deposited pooled ranks and equal-variance t statistic. The cleaned file had already removed missing, nonnumeric, and prompt-inconsistent estimates. That last rule conditions on an outcome after treatment, so it is reproduced historically but is not the live primary screen.
The printed Chicago row, d=1.79, t(5282)=65.00, N=5,284, already stood in the 2014 Table 2; the 2019 correction reprinted it unchanged while fixing a df and N typo in a different anchoring row. Many Labs also flattened the original two-step question, first higher or lower, then an estimate, into a single factual statement. The team's own randomized follow-up measured Chicago at d=0.50 in the question format against d=1.16 in the statement format, so format alone moves the magnitude. The live Houston prompt keeps the statement format.
Arm B: living readers
A split no one reader can see alone.
Opened 2026-08-21. Accepted rows: 0.
Waiting for the collective store.
DISPLAY CLOSED. The historical reproduction or analysis seal failed. Opening live results through untrusted analysis would turn a visible check into decoration.
No contributions yet
This arm opened 2026-08-21 with zero rows. At one row the page will show counts only. At 20 rows, with at least 5 in each prompt arm, it will open the pre-registered confidence sequence, effect estimate, distributions, and sensitivity views.
Counts only, while the arm is small
The thresholds of 20 total and 5 per arm are display choices, not a power calculation or stopping rule. The rows remain in arrival order.
Insufficient arm support
Twenty or more rows have arrived, but one prompt arm holds fewer than 5. A high-minus-low display fed by an empty or starved arm would restate assignment luck, not an effect, so the causal panels stay closed until both arms have the declared minimum.
Mean of the inverse-propensity score: unbiased for the accepted crowd's average effect over the random assignment, though with unequal realized arm counts it is not the simple difference of arm means. Direction was pre-registered as positive; no magnitude was promised.
Valid at every moment simultaneously if the prompt package shifts every contribution's expected response by the same amount, or if contributions behave as draws from one fixed population. If effects vary across readers or time, it is not a crowd-average interval and can even come up empty.
Ten to the power of the primary contrast times log10(8,000,001). With unequal arm counts this is not the observed ratio of the two arms' geometric means, and it is not a personal score.
An anytime test of the sharp null that the prompt package shifts no contribution's expected response. Under that null the chance of ever reaching the 20× threshold for alpha .05 is at most 5 percent. Evidence against that null, not protection against spam.
The confidence sequence calls the shared betting implementation on a disclosed grid of candidate means. Unlike a repeatedly viewed 1.96 standard-error interval, it pays for all present and future looks at once. Its guarantee assumes the condition is a fair independent draw, that contributions are not adversarially duplicated, and, for the interval reading, one common effect or a stable contributor population. The bankroll's sharp-null guarantee needs no such constancy.
Distribution, before any outcome filter
low prompthigh prompt
Twelve equal bins of Y = log10(estimate + 1) / log10(8,000,001). Binning is descriptive. The primary calculation uses every unbinned Y.
Sensitivity views, not the primary result
Ten-percent winsorized log-mean contrast, high minus low: . This caps each condition’s transformed tails before averaging and remains secondary.
| View | Low n | Low median | High n | High median |
|---|
Prior exposure is self-reported. The prompt-consistent view removes low responses at or below 200,000 and high responses at or above 4,000,000. Because that filter uses an outcome after randomization, it cannot replace the causal primary analysis.
Raw accepted audit panel
The frozen store exposes no integrity-flag or signed-token column. Therefore every syntactically accepted API row is both raw and primary here. Latency and exposure never exclude a row. The table shows every row returned, in arrival order.
| Day | Condition | Estimate | Seconds | Prior exposure |
|---|
Arm C: model completions
The same prompt, a different population.
Pending or absent
No usable model-arm file is present yet, or it contains no runs. Nothing is imputed. Each missing protocol remains missing.
Generated: . Rail: .
| Model | Protocol | Raw | Value | Refusal | Unparseable | Parsed median |
|---|
Registered response contrasts, kept separate by model
| Model | Houston isomorph | Famous-item panel |
|---|
Each contrast applies the same bounded log transform used for readers and reports high minus low plus the ratio of geometric means of answer + 1. No contrast pools different model names.
The registered parser trims ASCII whitespace only. A value must be the entire completion, digits only, no leading zero, and from 0 through 8,000,000. It does not remove commas, extract prose, retry, or ask another model to repair an answer. Refusal phrases are counted separately; everything else that fails is unparseable.
| Protocol | Model | Class | Raw completion |
|---|
Houston is the same structural prompt as the living arm. The African-UN protocols are a separate famous-item contamination panel. Model outputs are correlated samples from versioned systems, not independent people. Similar output shifts would not establish a shared cognitive mechanism.
The check
What is fixed, and what can fail.
Pre-registered analysis SHA-256
5b6cfd2f9ddecccd127488bc889345ba466d9f4d6cc52666fbbbc054af27086f
Checking the live bytes against this literal.
Off-page check
curl -s https://artwaste.land/strata/anchoring-rct/analysis.mjs | shasum -a 256
THE HISTORICAL ANCHORThe 1974 panel contains two published medians and no invented people. Arm A is visibly a 2014 replication. Its 5,284 cleaned observations are shipped as 518 lossless frequency cells and recompute n, medians, t, df, and d. A mismatch closes Arm B.
THE LIVE PRIMARYEvery accepted row enters Y = log10(estimate + 1) / log10(8,000,001). With high condition T, the bounded inverse-propensity score is 2TY minus 2(1 minus T)Y. Under fair assignment its average is unbiased for the accepted crowd's average effect, but it is not the observed difference of arm means when the realized arm counts differ. The direction high greater than low is registered; no effect size is promised.
ANYTIME VALID, FOR WHATStakes use only past rows, so two guarantees hold through every look. The bankroll is a level .05 anytime test of the sharp null that the prompt package shifts no contribution's expected response; that needs only fair assignment and bounded rows. The 95% sequence covers a common effect shared by every contribution, or the population average if contributions behave as draws from one fixed population. If the true effect varies across readers or time, the sequence tracks no single number: it is then not guaranteed to cover the running crowd average and can even become empty, which is evidence against constancy, not a crowd measurement. A repeatedly viewed fixed-n interval would offer none of this.
THE EMPTY STATEThe arm opened 2026-08-21. Zero rows say zero. From 1 through 19, the page shows counts only. At 20 total rows with at least 5 in each arm, the full registered display opens; with 20 or more rows but a starved arm, the page says insufficient arm support instead of showing a contrast. These are presentation thresholds, not a claim of adequate power.
ASSIGNMENT LIMITThe condition is an independent browser cryptographic bit and its local nonce is hashed before prompt display. The frozen API accepts only condition, estimate, latency, and exposure. It cannot issue or verify a signed assignment token, so the receipt is not described as one.
SPAM LIMITThe store limits an IP to two accepted rows per day, and a local marker discourages repeats. Neither proves one human per row. A determined Sybil attacker remains unsolved. Anytime-valid means safe through repeated looks, not safe through invented identities.
NO PERSONAL VERDICTOne person supplies only one potential outcome. The randomized contrast estimates a crowd-level effect of the whole prompt package. It cannot diagnose susceptibility, unconsciousness, irrationality, adjustment, accessibility, priming, or a stable trait.
CONCEPTUAL REPLICATIONThe live task changes the target, anchors, population, context, incentives, battery length, and statistic. It is neither an exact replication of 1974 nor Many Labs. Like Many Labs, and unlike the original two-step question, it asserts the anchor as a statement; the team's own follow-up found that format difference alone moved Chicago from d=0.50 to d=1.16. The fixed Houston item reduces iconic-task familiarity, but its Census fact may still be known.
HISTORICAL UNCERTAINTYThe 1974 demonstration’s N, group sizes, recruitment, exact wheel mechanics, truth value, and complete payoff manipulation are unreported in the article. A search found no public participant deposit as of 2026-08-21. That search cannot rule out private archival sheets.
MODEL LIMITEvery raw completion is retained and classified once as value, refusal, or unparseable. Model versions can change, samples can be correlated, and training contamination is generally unknown. Output similarity is not evidence of human-like cognition.
FREE CHOICESThe log transform, 0 to 8,000,000 bound, equal assignment probability, 1,001-point betting grid, alpha .05, 20-row display threshold, 5-row minimum per arm, 10% winsor sensitivity, exposure split, prompt-consistency sensitivity, 1 percent anchor-echo window, 12 histogram bins, and strict model parser were fixed in the sealed analysis or disclosed display code.
WHAT A NULL MEANSA wide sequence may mean little information. A sequence containing zero may reflect this self-selected, no-stakes population or this new item. It would not erase other prompt effects. A sequence excluding zero would still not establish one universal mechanism or magnitude.
Sources and audit trail
Where every anchor came from.
Tversky and Kahneman, Science, 1974Bibliographic record and the page 1128 summary: anchors 10 and 65, medians 25 and 45. No raw wheel-task rows or N are supplied.
Klein et al., Many Labs 1, 2014The 36-sample replication project and its four anchoring scenarios. Arm A is its Chicago item.
Klein, erratum, 2019Reprints the anchoring table. The Chicago row, d=1.79, t(5282)=65.00, N=5,284, already stood in the 2014 Table 2 and is unchanged there; the page compares those printed values with the shipped rows and recomputation.
Many Labs team, anchoring sentence format, OSFThe randomized follow-up comparing the original two-step question format with the flattened statement format: Chicago d=0.50 versus d=1.16.
Many Labs 1 datasets, OSFCC0 cleaned participant data. The downloaded 30,281,703-byte archive had SHA-256 c0d2ce04e3ea405baf5a2400e3d65bb71805e9d8a123ba7a71685f73bc4f3ff5 on 2026-08-21.
U.S. Census Bureau, 2010 decade factsThe fixed Houston city, Texas count of 2,099,451.
Howard et al., Annals of Statistics, 2021Time-uniform confidence sequences and empirical-Bernstein boundaries. The live calculation uses the house betting confidence-sequence implementation documented in the shared kit.
Röseler et al., OpAQ, 2022A large anchoring dataset and review that treats the basic effect as robust while naming theoretical disagreement and failed moderator or incidental-anchor findings.
The shipped page-data.json records extraction provenance and license beside each table. The cited records above are the public audit trail; the repository research record adds re-fetch commands and unresolved questions.