dictionary.py now holds one definition per column for both tables — name, dtype,
unit, description — and build.py renders it four ways: data-dictionary.csv,
docs/data-dictionary.md, the card's Fields table, and the Hub's dtype frontmatter.
None of those is hand-written any more. The descriptions were previously prose in
the README and the dtypes a separate dict in build.py, so the two could disagree
about the same column and nothing would notice.
check() fails the build when a builder emits a column the dictionary does not
document, or documents one it does not emit, order included. Verified by removing
a column from the dictionary and confirming build.py exits 1 rather than writing a
dataset whose documentation is wrong.
datasheet/datasheet.typ is a datasheet for the dataset rather than a reference to
the sizes: scope in and out, composition of both tables, how each family's numbers
were derived and why ALCs have no internal dimensions, known limitations, capture
provenance, and the…
Define the scope, and carry both unit systems on every row
Scope, stated in the card because it was only ever implicit: this dataset stops
at the unit. Dimensions, volumes, capacity, nesting, naming and standards
conformance for the tote itself; pallet builds, container loads and shelving
layouts are out. The one deliberate exception is eur1_pallet_per_layer, which is
the last figure that is purely a property of the box — everything past it depends
on build height and pallet standard, which are decisions rather than dimensions.
The card now also says what the thing is for: pallet, shelving and consolidation
calculations, and comparison on cost per usable litre.
Both unit systems on every row, in both tables, plus capacity in US gallons. The
comparison this dataset exists for is cross-market and one of the four vendors
designs in inches, so making half the readers convert before they can compare two
rows is where mistakes happen. listings.csv gains native_unit — which side is the
vendor's own figure and which is the conversion — because the de…
Eight pages, rendered from containers.csv and listings.csv rather than typed:
a family per section, each footprint's sizes with typical internal dimensions,
external volume, usable capacity and the vendor range, then the per-vendor
internal figures that range was computed from. Sources and naming pages at the
back.
container_stats.py now holds the per-vendor statistics that build.py and the
PDF both state as fact — same numbers, one definition. build.py output is
unchanged by that refactor, verified byte-for-byte.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Salesbridges (NL) publishes internal dimensions on 70 of 98 crate listings,
which more than doubles the evidence base and immediately shows the problem:
they quote 570x370 mm inside a 600x400 box where Plastic Box Shop quotes
555x355 and Yosibox 550x360. That is the top opening versus something near
the base, an 11 % spread in litres on the same box, and it was previously
invisible in the dataset.
- listings.csv: 52 -> 121 rows, four markets, EUR pricing added
- containers.csv: internal_dims_basis, internal_dims_vendors,
vendor_capacity_low_l, vendor_capacity_high_l on every row
- KLT rows 6429 and 4329 now carry vendor-published internal dimensions;
ALC rows still carry none, and guidance/internal-dimensions.md says why
- common_names expanded with regional and colloquial terms (Hudson box,
industrial tote, KLC, Eurobehälter, bac gerbable), with guidance/naming.md
separating what is evidenced here from what is trade usage
- scripts/capture_salesbridges.py: Lightspeed eCom sto…
Add listings table, raw captures, guidance, and full Euro height series
Second config (listings.csv): 52 individual vendor listings across UK, Israel
and US with prices, URLs and capture dates — the observations the typical
figures were generalised from, so the main table can be audited rather than
taken on trust. Raw trimmed captures under raw/.
Euro height series expanded to what was actually observed on sale: 18 heights
at 600x400, 9 at 400x300, 4 at 300x200. New listings_observed column records
how many real listings back each typical row, so interpolations into gaps in
the evidence are visible rather than implied.
guidance/eurobox-vs-alc-same-size.md answers the like-for-like question: a
sloped ALC gives up ~2.8 L (~4.4%) against a straight-walled Eurobox of the
same 60x40x34 external size, with the band and the reasoning.
Headline from the listings: ALC median usable ratio 71% vs Euro 77%,
consistent across three independent markets.
Declare column schema in the card; nesting ratio 0 rather than blank
The Hub inferred typical_nesting_ratio and notes as all-null from the CSV,
which breaks load_dataset for those fields. build.py now writes the
frontmatter schema in the same run that writes the data, and a
straight-walled box carries a nesting ratio of 0 rather than a blank.
Initial dataset: 35 container sizes across Euro, ALC and VDA KLT families
