What gets in

A document becomes a source if it states a value together with enough context to know what the value means. In practice that reduces to three questions: what was tested, at what temperature, and by which method. A figure that answers all three is transcribed. A figure that answers two of them is transcribed with a lowered confidence flag. A figure that answers one is left out.

The rule has a cost. Manufacturer literature is where most refractory numbers live, and most of it quotes a conductivity without naming either the method or the temperature. Applying the rule removed about a third of the values collected during the first six months of the project, and it is why 412 records took three years instead of three weekends.

The 38 sources

Each source carries an id, full bibliographic reference, year and a note on its limitations in sources.csv. Years run from 1971 to 2025; the median is 2004, which tells you something about where published refractory data actually sits.

Test methods referenced most often across the set are ISO 1893 for refractoriness under load, ISO 5013 for hot modulus of rupture, ASTM C20 for density and porosity, ASTM C133 for cold crushing strength and ASTM C201 for thermal conductivity. Castable records lean on the ISO 1927 series for preparation and curing. Where a source used a national method with no international equivalent, the method name is recorded in sources.csv and the record is capped at medium confidence.

Unit normalisation

Everything is converted to SI at transcription time, and the original figure is kept in the source note so a conversion mistake can be found later. The conversions that come up repeatedly:

Service temperatures are rounded to 10 °C because the underlying judgement is never finer than that, and a converted 2 642 °F printed as 1 450 °C would imply a precision nobody has. Chemistry is kept to one decimal, which is already generous for grades quoted as "about 70% alumina".

When two sources disagree

Disagreement is the normal case, not the exception. The resolution order is fixed, and it is applied per field rather than per record:

1. Prefer the value measured by a named standard method over one with no method.
2. Among named methods, prefer the one whose conditions match the field definition - a conductivity at 1 000 °C beats one at 800 °C extrapolated upward.
3. Among equals, prefer the more recent document.
4. If values still differ by more than 15% for a continuous property, keep the lower one, cite both sources, and set confidence to medium.
5. If they differ by more than 40%, the record keeps the field empty and the disagreement is written into the source note. Eleven records are in this state, all of them for hot modulus of rupture.

Nothing is averaged. An average of two incompatible tests is a number that was never measured, and it would be impossible to trace back to a document.

The confidence flag

high - 214 records

At least two independent sources, or one standard-method source, for every populated numeric field; no conversions deeper than a single unit change.

medium - 137 records

Single source, or a resolved conflict, or a conductivity marked calculated rather than measured.

low - 61 records

Older document with partial conditions, or a grade whose definition varies between sources, or a value we suspect but cannot check. Treat these as an indication of magnitude only.

The flag describes the record, not the family. A low-confidence microporous panel sits next to a high-confidence firebrick in the same insulating family, and filtering by family will mix them unless you also filter on confidence.

How full the columns actually are

Known gaps

Written down so nobody has to discover them the hard way:

• Hot strength is barely covered. Only 29% of records have a hot modulus of rupture, and the field does not even carry the test temperature. Fixing this properly means a schema change, which is why it is still provisional.
• Thermal shock counts are weakly comparable. Three protocols, one ordinal outcome, and no information about specimen size. Comparing cycles within one protocol is defensible; across protocols it is not.
• Conductivity above 1 200 °C is mostly absent. Sources stop where the furnaces of their era stopped. For high-temperature behaviour you will have to go to the primary literature yourself.
• Carbon-bearing grades are documented for reducing conditions only. The oxidation limit in air, which is what decides whether a lining survives a shutdown, exists for just 11% of records.
• Geographic skew. Sources are predominantly European and Japanese. Grades common in North American or Indian practice are under-represented, and some are missing entirely.
• Nothing on creep, abrasion or alkali attack. All three matter in service; none had enough comparable published data to normalise, so no field was created rather than creating one that would be 8% full.
• No availability or cost data, and none planned. It changes faster than anyone could maintain it here, and it would date the whole release.
• Grade drift is invisible. A record describes a grade as published in a given year. The product sold under that name today may be different, and the dataset has no way to show that.

Review

Every change is reviewed by the maintainer who did not make it. The review checks three things: that the cited source really contains the value, that the conversion arithmetic is right, and that the confidence flag follows the rules above rather than the reviewer's optimism. Reviews are recorded against the release, so each entry in the changelog can be traced to the records it touched.

Two of us do this in evenings and the occasional quiet week, which sets the pace honestly: a correction with a clean citation is usually in the next minor release, and a new family takes a season.