Skip to content

Observation · 08 / Illustrated journey

How much can a finite record establish?

Compare a stated probability with a synthetic outcome count, keep the sampling assumptions visible, and inspect a confidence interval that stays within the probability range.

A synthetic observed fraction is compared with a supplied prediction and a Wilson interval. Increasing the sample size narrows sampling uncertainty under the independent-trial model.
A synthetic observed fraction is compared with a supplied prediction and a Wilson interval. Increasing the sample size narrows sampling uncertainty under the independent-trial model. Download SVG
01

A fit and a prediction answer different questions

A physical proposal becomes testable when it specifies both what should be measured and what alternatives would produce different records. Reproducing a formula that was inserted into a construction is a consistency check. A distinguishing test needs an independently fixed prediction and a controlled comparison. The book makes that separation central to its research programme. [O04]

The laboratory supplies a probability p₀ and a synthetic count k out of n independent trials. It displays the observed fraction k/n and a two-sided 95% Wilson score interval. No actual experiment is connected to the page, and entering a count does not establish that its sampling assumptions hold.

The uncertainty calculation concerns a constant binary probability. It does not include detector drift, unknown efficiency, correlated trials, model selection, or repeated searching across many settings. Those effects require additional measurements or a more appropriate likelihood.

02

Separate disagreement from sample size

Change the number of trials n, the requested observed fraction f, and the supplied prediction p₀. The lab converts f to a valid integer count k=round(nf), then uses the actual estimate k/n. This matters for small samples: only fractions compatible with integer counts can be observed. Compare that estimate with the interval before interpreting a visible difference.

Interactive calculation

Change the inputs to compare the model outputs.

Enable JavaScript for adjustable values. The static figure and worked calculation remain available.

For the synthetic example n=400 and f=0.55, k=220 and the Wilson interval is approximately [0.5010,0.5980]. A supplied probability 0.50 lies just outside this interval; 0.54 lies inside. These are outcomes of this one fixed comparison, not a verdict on an entire physical theory. Away from probability boundaries, interval width contracts approximately as n⁻¹ᐟ².

03

Invert a score comparison

p̂=k/n,   z=1.959963984540054,   d=1+z²/n
c=(p̂+z²/(2n))/d,   h=(z/d)√[p̂(1−p̂)/n+z²/(4n²)]
Wilson interval = [c−h,c+h]

The formula comes from solving the score-test inequality for the unknown probability. Unlike the elementary symmetric interval p̂±z√[p̂(1−p̂)/n], it remains inside [0,1] and does not collapse to zero width merely because the observed count is zero or n. The NIST handbook gives the expression and distinguishes it from the adjusted Wald interval. [O05]

For k=0, the upper limit simplifies to z²/(n+z²), while the lower limit is zero. Thus zero observed successes does not establish a zero underlying probability. Likewise, k=n does not prove certainty.

A confidence procedure is interpreted through repeated sampling: its intervals have an intended coverage level under the assumed model. For a realized interval and fixed unknown p, it is not automatically a posterior probability statement about p. Discrete counts mean the Wilson procedure's actual coverage is approximate, not exactly 95% for every n and p.

04

Build a comparison that can survive repetition

Before collecting data, identify the outcomes, exposure, inclusion rule, calibrated apparatus model, parameter estimates, and reserved comparison. If p₀ was tuned to the same k, agreement is partly built into the exercise. Reserving observations or predicting a second setting makes that dependence testable.

A systematic bias does not shrink just because n grows. Independent batches, settings in a randomized order, detector reversals, and calibration checks can distinguish a model discrepancy from an apparatus effect. These are design choices for an actual study; the slider experiment supplies none of that empirical evidence.

For SMM/DTT, the next scientific step is an explicit source model whose retained-record statistics differ from a stated baseline after its scales and nuisance parameters have been fixed. The published reconstruction gives mathematical constraints that such a model must satisfy. The website's laboratories make assumptions inspectable and calculations reproducible; they do not convert a supplied standard model into a new experimental prediction. [P01] [B16]

Return to Emergence for generated records, Matter for persistent physical carriers, or the source and model register for the complete conventions used here.

SMONAD

Share and cite

Share this page, or keep a clear reference to the work you are reading.

More platforms
Cite this page

Cite the website explanation

Use this reference for the explanation on this page. “n.d.” means its first publication date is not specified; the update date is recorded separately.