Twelve manually transcribed Table 1 rows must recompute .693 correct, .739 confidence, and .046 difference. If any fails, the living display stays closed. The 43%, 49%, 37%, and 45% range figures are labelled published summaries because no underlying range responses were found.
Arm B, the living
Draw eight fences around what you do not know.
A subjective 90% range is not a sampling-theory confidence interval. It is your statement that, over many questions like these, you intend about nine truths in ten to land between your endpoints. With eight fixed items, exactly 90% is impossible: 7 of 8 is 87.5% and 8 of 8 is 100%.
Checking the historical anchor before opening the instrument.
Before you begin: We cannot stop you searching. Please do not look these up if you want the intended memory-and-judgment task. The page cannot prevent model help, repeat visits, answer sharing, or scripted submissions.
The counter that can be watched
Every look is already paid for.
A 95% confidence sequence is valid at every moment simultaneously, under its bounded conditional-mean assumptions. A familiar fixed-sample interval such as an estimate plus or minus 1.96 standard errors would quietly lose its advertised guarantee when visitors peek repeatedly. This display uses the shared betting confidence sequence and shows its grid. What accumulates here is a client-reported coverage mask: the store keeps no endpoints, so a deliberately enormous range counts as covered and cannot be told apart from an informative one, and no format-informativeness claim can be tested on this wire.
Waiting for the anchor check.
Arm A, the published record
A table first. The famous percentage second.
Klayman and colleagues' Experiment 1 printed twelve domain aggregates. Those cells can be transcribed and averaged. Their Experiment 3 printed only summary percentages, so 43% remains a cited benchmark, never a fabricated row set.
Loading the transcribed cells.
| Domain, Table 1 p. 226 | Correct | Confidence | Printed difference |
|---|
Different cohorts: Experiment 1's two-choice result and Experiment 3's subjective-range result did not come from the same people. Experiment 3 recruited 32 new University of Chicago students. Its method also leaves an unresolved denominator: it refers back to a 120-question procedure while its printed instruction says 150.
Arm C, the machines
The same facts, raw answers kept.
Tool-disabled model snapshots receive either one direct-range prompt or independent lower and upper prompts. Parsing is strict JSON with no repair, retry, or silent dropping. This tests a dated model and prompt on eight possibly memorized facts. It does not test sentience, a stable trait, or an uncontaminated mind.
Model sampling is pending. Nothing is being imputed while the file is absent.
The check
What can fail here.
The analysis file was fixed before this arm opened. Its literal SHA-256 is visible here and checked against the live bytes on every load.
Sealed analysis SHA-256
078967839da154a48f1f54dbb4c5771d5b3025b45915b3b789c7a0a307f37fec
Checking the live analysis bytes.
No original Experiment 3 item responses, endpoints, full domain table, exact question list, analysis code, or resolved trial denominator was located as of 2026-08-21. No pseudo-trials are manufactured from 43%.
The frozen store has no signed token or server assignment field. This page randomizes condition in the browser and stores only the assigned integer. A visitor can alter it. The difference is an exploratory randomized invitation, not a tamper-proof causal estimate.
The frozen four-integer payload permits screening only on item-set version and elapsed time from 10 through 900 seconds. It cannot verify endpoint completeness after submission, assignment signatures, or one person. Exclusions never use correctness.
The confidence sequence tolerates arbitrary viewing and stopping under a bounded conditional-mean model. Repeat submitters, dependence, answer sharing, and organized attacks can break that interpretation. The 1% and 5% bands are contamination sensitivity calculations, not estimates of spam.
Eight editorially selected facts are not a random sample of questions. Per-item rates are descriptive. Neither 7 of 8 nor any other personal score diagnoses a stable tendency, and the crowd does not estimate humanity.
Coverage rewards very wide ranges. The page shows your widths locally, but the store does not retain endpoints, so crowd width and lower-tail versus upper-tail misses cannot be computed. No claim that one format improves informativeness is made.
The keys are institutional reference values, not metaphysical exact truths. Denali can be remeasured, Mercury parameters refined, museum dimensions revised, and census totals depend on inclusion rules. This item set never changes under existing rows.
The source pages were read, but byte snapshots were not archived. Their snapshot-hash fields are null and this weakens reproducibility. Retrieval dates, exact definitions, conversions, and URLs are shipped instead.
Every raw completion becomes value, refusal, or unparseable, and those counts must sum to all runs. Split calls pair by model and arrival index because the supplied result schema has no replicate identifier. Unmatched or reversed pairs remain visible failures.
Later human work already shows elicitation matters, and 2025 model studies already report numerical-interval undercoverage. This page can compare a version-pinned public crowd with model snapshots. It cannot claim to discover either phenomenon.
No original dataset or data licence was found. The article PDF is copyrighted. The twelve factual aggregate rows are a minimal transcription, with editorial and legal review still required before publication.
Sources and edges
Where every number came from.
Klayman, Soll, González-Vallejo, and Barlas, 1999Experiment 1 Table 1 on p. 226, Experiment 3 on pp. 238-241, and the 47-point summary on p. 242. The page uses only transcribed aggregates and labelled summaries.
Soll and Klayman, 2004Direct versus separately elicited endpoints, the prior basis for the format comparison.
Teigen and Jørgensen, 2005Further elicitation experiments. Its 52.5% figure is stated confidence, not coverage; the paired hit rate was 23.4%.
Howard, Ramdas, McAuliffe, and Sekhon, 2021Time-uniform confidence sequence foundations. The deployed calculation uses the site's disclosed shared betting implementation.
Bahaj and colleagues, 2025, preprintPrior LLM numerical-interval calibration work. It is not directly comparable to these eight items.
FermiEval, 2025/2026 preprintA closer model benchmark with interval scoring and conformal adjustment. This page makes no novelty claim over it.
The eight truth-source links appear beside the revealed answers after you score. Their complete audit fields live in page-data.json. The research README lists the searches, data boundary, licence uncertainty, and re-check instructions.