Research / Rule boundaries

When a centimeter changes the label: boundary sensitivity in a body type calculator

A body type calculator turns four measurements into a label by comparing ratios with fixed thresholds. Near those thresholds, a small change in one measurement can change the label. This analysis scans Bodytype's v2 rule space to measure how thin those boundaries are - not how often real people change labels.

By Bodytype.org · Updated 17 November 2026 · About 12 minutes
Synthetic rule-space analysis · Algorithm v2 / dataset 1.0

Editorial note: this page is published after an internal methods review of its denominators, rates and wording. It is not external peer review, statistical certification or evidence that the calculator is accurate.

Key findings

5 findings · 2026-11-16 run

Five results to keep in view

  1. 23.6% of tested points changed label under a ±1 cm shift.

    Among 669,858 tested rule-space points, 158,352 (23.6%) changed label under at least one single-field ±1 cm shift. The shares were 76,113 (11.4%) at ±0.5 cm and 307,610 (45.9%) at ±2 cm.

    Table 1Summary JSON

  2. Nearly one in four tested points had a stability radius below 1 cm.

    154,674 points (23.1%) were below 1 cm and 421,650 (63.0%) were below 3 cm. Another 43,715 points (6.5%) sat exactly on a threshold; no point had an infinite radius.

    Stability chartLimitations

  3. The bust-to-hip thresholds were crossed most often.

    bustToHip:1.05 was crossed in 123,702 of 324,100 ±1 cm flip events (38.2%), followed by bustToHip:0.95 (119,034; 36.7%). 12,096 events crossed more than one threshold, so these shares overlap.

    ResultsHow to read

  4. Rounding to 1 cm or 0.1 in changed more labels than rounding to 0.5 cm.

    Rounding inputs to 1 cm changed 20,095 labels (3.0%); rounding to 0.1 in changed 25,820 (3.9%); rounding to 0.5 cm changed 1,449 (0.2%).

    Results

  5. These are rule-space properties, not population estimates.

    The grid is synthetic and deterministic. The label shares below describe the tested grid, not how common a body type is, and the rates are not probabilities that a real user will change labels.

    How to read the numbersLimitations

What this measures

How thin the rule boundaries are on a synthetic grid.

The scan changes one measurement at a time and records whether the v2 label changes. It also records how far each tested point sits from the nearest threshold.

What this does not measure

Accuracy, population shares or real measurement error.

No person was measured. The rates are shares of tested rule-space points, not probabilities that a user will see a different label.

01 / The rule

Why a centimeter can matter

Bodytype v2 does not compare raw centimetres. It turns four measurements into three ratios: bust to hips, waist to hips, and high hip to hips. Each ratio is compared with fixed thresholds, and the first matching rule wins. The rule order and constants are the same ones the calculator uses at runtime.

That design is simple and reproducible, but it has a geometric consequence. Near a threshold, a small change in one input can move a ratio across the line. A 1 cm change in hips changes all three ratios at once; a 1 cm change in bust, waist or high hip changes only one. This is why hips has the highest field-level flip rate in the results. It is not evidence that hip measurements are less reliable.

Table A. The five v2 ratio thresholds
ThresholdRule role
Bust / hips 0.95Lower bound of the balanced bust-to-hip range; below it the rule can return Pear.
Bust / hips 1.05Upper bound of the balanced bust-to-hip range; above it the rule can return Inverted triangle.
Waist / hips 0.80Hourglass requires a waist-to-hip ratio at or below this value.
Waist / hips 0.88Apple requires a waist-to-hip ratio at or above this value.
High hip / hips 0.95Hourglass requires a high-hip-to-hip ratio at or below this value.

These are Bodytype's styling rules, not an ASTM, ISO, GB/T or medical classification. The values are read from the current rule implementation.

02 / The scan

What we scanned

The main grid is a synthetic rule space, not a sample of people. It fixes seven hip sizes and a dense grid of the three ratios, then keeps only the ordered combinations where waist < high hip < hips. Those are the combinations the calculator treats as a coherent set of four circumference inputs.

Hips80–140 cm · 7 levels
Bust / hips0.900–1.100 · 41 levels
Waist / hips0.650–0.950 · 61 levels
High hip / hips0.700–1.000 · 61 levels
Ordered constraintwaist < high hip < hips
Main grid669,858 points

Ratios are generated from integer thousandths to avoid step drift. The 198 warning-case points are a separate supplementary set and are not in this denominator.

For every main-grid point, the scan changes one field at a time by ±0.5 cm, ±1 cm and ±2 cm. A change is a flip when the v2 label after the shift differs from the base label. The scan also rounds all four inputs to 0.5 cm, 1 cm and 0.1 in, and records the smallest single-field shift that reaches a different label — the stability radius.

The scan calls the project's own classifyId, findNearestShift and rule constants. It does not copy the classification logic or define a second set of thresholds.

03 / Reading the numbers

How to read the numbers

The short version

A 23.6% flip rate means that share of tested rule-space points changed label. It does not mean that share of people, and it is not a probability for any one person.

  • Every denominator is a count of tested points, not a count of people. The grid is not weighted to any population.
  • The rates are deterministic shares, not estimates from a random sample. No confidence interval is reported because there is no sampling distribution to describe.
  • Field columns can overlap. One tested point can flip under more than one field, so the four field rates cannot be added to get the any-field rate.
  • Threshold-crossing shares can also overlap. 12,096 ±1 cm flip events crossed more than one threshold, so the five shares sum to more than 100%.
  • The 6.5% boundary share is an alignment between the 0.005 ratio grid and the thresholds. It is not a population share of people who sit exactly on a boundary.

04 / Results

Results

The numbers below use the 2026-11-16 run. All counts and rates are in the summary JSON; the tables and charts are a reading aid, not a second source of truth.

4.1 Overall flip rates

Single-field shift · label changed

Table 1. Single-field shift and label change
ShiftPoints that changedShare
±0.5 cm76,113 / 669,85811.4%
±1 cm158,352 / 669,85823.6%
±2 cm307,610 / 669,85845.9%

A point is counted once per shift size if at least one of the four fields changes the label in either direction. Positive and negative ±1 cm shifts were nearly identical, so the table pools directions.

4.2 Which field moves the label

Share of all tested points

Table 2. Field-level flip rate by shift size
Field±0.5 cm±1 cm±2 cm
Bust8.4%18.1%36.9%
Waist2.0%4.3%8.8%
Hips10.5%22.7%43.5%
High hip1.2%2.5%5.2%

A point can be changed by more than one field; the columns cannot be added. Hips changes all three ratios at once, which is why its rate is highest.

4.3 How close are results to a boundary?

Stability radius · cumulative

Figure 1share at or below radiusStability radius, cumulative share of tested points
  • 0 cm6.5%
  • ≤ 0.5 cm11.9%
  • ≤ 1 cm24.1%
  • ≤ 2 cm46.9%
  • ≤ 3 cm65.0%
  • ≤ 4 cm81.4%
  • ≤ 5 cm92.0%
  • ≤ 6 cm97.8%
  • ≤ 7 cm100.0%

Boundary points (radius 0) are included in the below-1 cm tier and listed separately. The full 180-row curve is in the stability CDF CSV.

4.4 Hip size

±1 cm flip rate by level

Figure 2seven tested hip sizes±1 cm flip rate by hip-size level
  • 80 cm30.9%
  • 90 cm30.9%
  • 100 cm27.5%
  • 110 cm19.1%
  • 120 cm19.1%
  • 130 cm19.1%
  • 140 cm19.1%

Each level has the same number of tested points. The same 1 cm shift is a larger share of a smaller circumference; this is not a population conclusion. Data: flip-by-hips CSV.

4.5 Where labels move

±1 cm single-field shifts

Figure 3. Label transitions under ±1 cm single-field shifts.

Hourglass

198,450 points
  • Inverted triangle22.4%
  • Pear21.8%
  • Rectangle18.8%

Pear

163,380 points
  • Hourglass15.0%
  • Rectangle9.2%
  • Apple3.0%

Inverted triangle

163,380 points
  • Hourglass17.0%
  • Rectangle9.0%
  • Apple2.6%

Apple

35,280 points
  • Rectangle39.0%
  • Inverted triangle22.4%
  • Pear21.8%

Rectangle

109,368 points
  • Inverted triangle22.4%
  • Pear21.8%
  • Hourglass19.8%
  • Apple7.9%

Bar length is the share of the base label's tested points. The largest single transition is Hourglass → Inverted triangle (44,550 events); the full matrix is in the transition matrix CSV.

4.6 Documented rule boundaries

Rule choices, kept visible

Two v2 decisions, kept visible

Both are deliberate v2 choices. They are reported so the numbers are not read as cleaner than they are; neither changes how the main results are read.

  • Unit round-trip: v2 compares ratios exactly and does not add a tolerance. Exact cm→in→cm conversion changed 3,417 labels out of 669,858 (0.5%). All 3,417 sit on boundary points — 7.8% of the 43,715 boundary points. A ratio of 0.95 can become 0.9499999999999998 and flip Hourglass to Pear. Adding a tolerance would be a new rule version, not a correction to v2.
  • High hip = hips: v2 treats equality as an accepted input and does not warn. In the supplementary set of 198 points, 162 produce a position-check warning; the 36 equality points do not, because only a strictly larger high hip triggers the check.

05 / For users

What this means for users

Close to a boundary, the useful response is a second reading, not a search for a more definitive label. The calculator already names the neighbouring pattern when the stability radius is small. That is a prompt to check the tape, not a probability that the other label is correct.

  1. Recheck the tape before treating the label as settled.

    Keep the tape level, stand without pulling the tape tight, and take the reading twice.

  2. Read the neighbouring pattern as a check, not a probability.

    The result names labels reachable within a small shift. It does not say which label is more likely to be right.

  3. Use the label as a styling reference.

    It is not a medical diagnosis, health assessment or an ideal-body standard.

06 / Limits of meaning

What this does not mean

  • It is not an accuracy study. No person was measured, and no label was compared with an expert or a second calculator.
  • It is not a population sample. The grid is synthetic and not weighted to any population, so the label shares do not say how common a body type is.
  • It is not a probability. The rates are deterministic shares of tested points, not estimates of how often a real user will see a different label.
  • It is not a medical or health assessment. The thresholds are Bodytype’s styling rules, not WHO, ASTM, ISO or clinical boundaries.

The boundary is not a verdict

A label changing near a threshold is a property of the rule, not a statement about the body. The result is a styling reference, not a diagnosis.

07 / Method and sources

Method, sources and limits

The scan uses the v2 rule functions and constants from the site repository. It records the base label, the stability radius, the labels reachable at that radius, the nearest threshold, three perturbation sizes, three rounding paths, and an exact centimetre-to-inch-to-centimetre round trip. The run checks unique IDs, the expected grid count, value ranges, deterministic classification, tier consistency, valid neighbours, transition totals, and that every recorded flip crosses at least one threshold.

The grid is fixed before the run. The analysis reports counts and shares with their denominators, not confidence intervals. All numbers on this page come from the generated summary; the method snapshot records the exact grid, perturbations, checks and SHA-256 values.

  • ASTM D5219-25Official listing

    Defines measurement terms and the default procedure, not the five labels or the thresholds.

  • Bodytype v2 rule implementationInternal

    Generated the method snapshot; checked in the 17 November 2026 internal methods review.

  • Published releaseArchived

    GitHub v1.0.0 · Hugging Face revision 2630867 · Zenodo 10.5281/zenodo.24283466.

  • Public measurement wordingSite guide

    Keeps the body-site names aligned with the terminology readers see in the calculator.

Limitations

  • The grid is synthetic. It does not represent any population, clothing size system or real measurement distribution.
  • The rates depend on the grid spacing and the v2 thresholds. They do not transfer to another rule or dataset.
  • No random sampling was used, so no confidence interval is reported.
  • Unit round-trip and high hip = hips are documented v2 decisions, not claims about measurement accuracy.

Last updated 17 November 2026. The date changes only when the content is revised, not when the site is deployed or the scan is rerun.

08 / Data and citation

Data and citation

The files below are the public snapshot. The full 669,858-row grid is available as a 3.2 MB parquet file on Hugging Face. The analysis code is on GitHub, and the release is archived on Zenodo under DOI 10.5281/zenodo.24283466.

Summary JSON

All counts, rates and denominators used on this page.

Download

Grid sample

A small CSV sample for checking the schema.

Download

Method snapshot

Grid, perturbations, checks and SHA-256 values.

Read

Code

Generation script, schema, sample and summary · GitHub v1.0.0.

Open repository

Licence and attribution

Code: MIT. Synthetic data: CC BY 4.0. Any use of the code, data, tables or figures must credit Bodytype.org and link to https://bodytype.org.

The code licence and the data licence are separate. The site article is the narrative canonical; the external code and data records link back here.

Cite this page

A stable reference for the boundary scan

Bodytype.org. (2026). When a centimeter changes the label: boundary sensitivity in a body type calculator (algorithm v2, dataset 1.0). https://doi.org/10.5281/zenodo.24283466

The code repository, dataset mirror and DOI resolve as of 2026-11-16. The site page and the archived release describe the same snapshot. This page is not a peer-reviewed publication. Boundary points: 43,715. ±1 cm events: 324,100.