12 calibers, 9 axes, and every hole left open

What this view cannot draw

Not one of the 12,269 designs in this corpus carries all 9 of these properties. The most complete record on the site carries 8. That is not a defect in the view — it is the single most useful thing the view says, and you can read it off the plot in one glance the first time a ribbon breaks. This page is deliberately not a corpus overview: 12,269 polylines would be solid ink, and 12,269 is a fifteenth of what this corpus will hold when the current ingestion lands. It is a comparison instrument, and the cap of 12 is the form working rather than the form giving up.

Designs by how many of the 9 axes they carry. The column at 9 is empty, drawn as a dashed rule rather than as a bar, because an unknown must never read as a small quantity.

Reading the axes left to right: 11,084 designs carry the first 2 · 9,743 designs carry the first 3 · 8,918 designs carry the first 4 · 8,357 designs carry the first 5 · 7,184 designs carry the first 6 · 6,637 designs carry the first 7 · 3,669 designs carry the first 8 · 0 designs carry the first 9. Every one of those drop-offs is a ribbon breaking, and the drop is steepest at the last axis because it is the least documented property in the vocabulary.

The 300 designs this view can put nothing on the picture for

They carry none of the 9 axes. They are not filtered out: add one and it appears as a numbered chip, a labelled row, a full row of hollow ticks in the gutter, and a table row of dashes. What it does not get is a line, because there is nothing to draw one through. The first 12 by descendant reach: Venus 188 / Valjoux 7730 · Eterna 1427 · Rolex 1530 · Cartier 1847 MC · Baume & Mercier Baumatic BM12-1975A · ETA 2890-2 · Lemania 5100 · Rolex 3085 · Venus 175 · Ракета 2609.НА · Credor 6830 · Grand Seiko 9F62. That list is a warning about the corpus, not about the calibers: “Venus 188 / Valjoux 7730” is not an obscure movement.

And the gaps are not random

Mean number of these 9 axes carried, by maker, over the 71 makers with at least 5 designs attributed. Best: Enicar 7.66 over 67 · Zodiac 7.57 over 14 · Mido 7.53 over 36. Worst: Tianjin Sea-Gull Watch Group 0 over 7 (7 with nothing at all) · Luch 1.05 over 43 (33 with nothing at all) · Molnija 1.17 over 6 (4 with nothing at all). The 5,770 designs with no maker sourced average 6.7. A comparison that happens to pick from the top of that list will look complete; one that picks from the bottom will look broken. Both pictures are true, and neither is a statement about the movements.

How a value gets chosen when the corpus asserts more than one

Where a design asserts several values for one property this view never averages and never chooses: values in a dispute group are drawn as ruby ticks and the ribbon breaks between them, and values that merely differ in scope (the 17/21-jewel case) are drawn as a fork through every asserted value. rank='preferred' would supersede this behaviour; the corpus currently has no preferred rows.

And how one never gets chosen: there is no code path in this view that joins two axes across a missing value. A gap is emitted as a gap in components/viz/compare-geometry.ts, and tools/layout_compare.py re-derives every carrier count from SQL and refuses to write the payload if it disagrees with what it is about to publish.

The axes, and why each one is on the scale it is on

1 Jewels · linear

corpus range includes 0, so the axis is linear. Extremes: Dugena 819 (Otero 27) at 17,721 · 4002 anonymous at 0.

2 Frequency · log · vph

corpus range 120–5,184,000 spans a factor of 43200; a linear axis would be governed by its largest value alone. Extremes: Omega 1220 at 5,184,000 · Jaeger-LeCoultre 560 (Atmos) at 120.

3 Movement height · log · mm

corpus range 0.98–15.1 spans a factor of 15; a linear axis would be governed by its largest value alone. Extremes: Ganthony 19''' at 15.1 · ETA-ESA 210.001 at 0.98.

4 Winding mode · enum

enum property; 2 of 4 vocabulary levels occur in the corpus.

5 Power reserve (typical) · log · h

corpus range 3–744 spans a factor of 248; a linear axis would be governed by its largest value alone. Extremes: A. Lange & Söhne L034.1 at 744 · Hanhart 155 at 3.

6 Movement diameter · linear · mm

corpus range 9–54 spans a factor of 6.0; linear keeps the units readable. Extremes: Slava 5498 at 54 · Ronda 1032 at 9.

7 Seconds display · enum

enum property; 4 of 4 vocabulary levels occur in the corpus.

8 Shock protection · enum

enum property; 11 of 12 vocabulary levels occur in the corpus.

9 Total components · log

corpus range 119–1,278 spans a factor of 11; a linear axis would be governed by its largest value alone. Extremes: Patek Philippe Calibre 89 at 1,278 · Jaeger-LeCoultre 849 at 119.

Two grains in one table, and they are labelled rather than blended. The first two columns count designs; the next three count asserted values. They differ wherever a design asserts more than one value for one property — the 17/21-jewel case — so a value count can and does exceed the count of designs carrying that property. Nothing is double-counted; the two columns are answers to different questions, and adding one to the other is meaningless.

Axes are ordered by coverage descending (n_designs), ties broken by property id, computed at build time, so the densest axes lead and the ribbons are established before the sparse ones thin them. Round round/2026-09-08 · corpus 70c0140ebad3f9a6 · built 2026-09-08T17:57:55 · every number on this page comes from tools/layout_compare.py.

A quantity with at most 16 distinct values in the whole corpus is drawn as an ordinal band rather than a continuous line, because a continuous axis would invent positions no escapement can physically occupy. A quantity whose corpus range spans a factor of 10 or more, with a positive floor, gets a log axis so one true outlier cannot govern the whole scale. Both thresholds live in the build tool, not in the page.

There is no time axis here on purpose. 8,569 of 12,269 designs have no introduction year, and 9,750 touch no lineage edge. The universe already places a dated caliber on a real era axis and holds the undated ones in its reservoir band rather than interpolating them; repeating that here would add a tenth axis on which most ribbons break for a reason this view has nothing new to say about.

Assemble a set

Sets to start from

Each of these is chosen by a rule at build time, not curated by hand, so the set changes with the corpus rather than going stale. The rule is printed beside it.

The comparison as a table

Everything the plot draws, in text, with the status of each value: ● a value no source contradicts · ◆ contested · — no value in the corpus. Ten columns do not fit a phone, so this table scrolls sideways inside its own box; the page never does.