Bound values come from real distributions, not imagined extremes

What it is

The shortest and longest strings used to stress a slot must be pointable in real data. Distribution-derived bounds take percentiles from that field’s production distribution — or from a comparable product’s — and turn those records into samples, instead of inventing a sentence that “looks long enough.” Imagined extrema usually get the shape wrong: a run of repeated letters, a cartoon German compound, a Chinese name that is always two characters, versus the real tail (a spaced compound family name, a ticket title with an ID, a filename with a path). Those are different geometries.

This answers where the two end values come from, not whether the end frames must be shipped. Without a distribution, the frames can look complete and still fail the live tail.

Why it happens

Length on a canvas is estimated from the samples a person has seen and from working memory of “exaggerated enough.” That estimate is kind to the mode and blind to the tail. Fields almost never break at the mode. They break at the shape of the 95th or 99th percentile — not the same sentence made longer, but another structure: an ID in the middle, prefixes, runs of spaces, a URL that will not wrap. A container built for a cartoon long word reserves for a case that never arrives, then fails on the first real long ticket.

Percentiles have a direction. Designing to the maximum treats a rare paste attack as daily furniture and turns the card into a monument. Designing to the median as a ceiling dumps an everyday tail on implementation to crop. The package has to say which slice it is taking: display through P99, and beyond P99 follow that type’s overflow policy. Without that sentence, the long sample is a pretty picture unhooked from production.

Where it stops holding

A brand-new product has no production data: borrow a distribution from the same kind of object (names from an existing account system, ticket titles from support history) and re-take percentiles once live data exists; do not freeze the borrow as a lifetime ceiling. Lengths nailed by a standard — card numbers, national IDs, fixed-width codes — are specified lengths, not distributions; sample the spec length and skip the histogram. Adversarial paste (a novel dumped into a nickname) should be stopped at input and must not enter the “real distribution” that sizes everyday layout. Internal tools whose names all come from one employee-id rule have a narrow distribution; sizing them to a consumer social tail wastes density.

Applying it

• For each critical text slot, pull field length (characters and rendered width) from the last release cycle. Report at least median, P95, P99, and max. Turn real P95 and P99 records into samples in the package.
• Label each frame with the percentile it represents, and state which overflow path applies beyond P99, so the maximum is not silently treated as “must show in full.”
• Do not use a Latin tongue-twister or repeated glyphs as the long sample unless production actually looks like that.
• Paste a live P99 record into the built UI and check that it lands inside the geometry the package drew. If the “extreme” in the file is shorter than P95, or an order of magnitude wilder than the historical max, the bounds did not come from the distribution.

Related

• Same group: R2.05.1 Handoff must show both extreme-long and extreme-short content · R2.05.2 Zero, one, and many each need their own defined layout · R2.05.3 Whatever is left unspecified will be invented by the implementer · R2.05.5 Overflow behavior must be specified separately by content type · R2.05.6 Stacked edge cases need their own frames
• Adjacent: R2.01 Annotation and asset export · S1.01 Text Expansion and Layout Resilience
• Search terms: distribution-derived bounds · percentile content length · empirical extrema