Trust artefact
Corrections log
Dated, numbered, oldest at the top. Nothing here is ever deleted or quietly rewritten — an entry that turns out to be wrong gets a later entry, not an edit.
The log opened on 2026-08-10, with the site. It runs in notebook order: the first entry is at the top and new entries are appended at the bottom. Ids run RP-COR-<year>-<nn>, sequential, never reused.
Three kinds of entry. A limitation is something our own work cannot support — filed by us, about us, ideally before anyone else has to point it out. A correction is a figure or a claim we published that was wrong: what it was, what it is now, and how it got past us. A clarification is a figure that stands but whose provenance a reader could reasonably misread.
An entry stays open until the thing it describes is resolved, at which point it gets a closing entry with its own date. Nothing is closed by deletion.
| Entry | Logged | Kind | Subject | Status |
|---|---|---|---|---|
| RP-COR-2026-01 | Limitation | What the first run will not be able to support | open |
What the first run will not be able to support
Filed before the first food-energy run publishes, deliberately. We hold other people's numbers to a standard of cross-lab replication, and it would be dishonest to let our own first result arrive without saying plainly what it cannot carry.
- One operator, one rig. Every plate in the first run is portioned and weighed by the same person on the same balance. Portioning is a judgement call at the margins, and we have no inter-operator check yet. Until we run one, treat the spread between products as more trustworthy than any single product's absolute figure.
- Fewer than 200 samples. At that size the confidence interval around a mean absolute percentage error is wide enough to swallow small gaps. If two products land within about a percentage point of each other, we are not entitled to rank them and neither is anyone quoting us.
- Restaurant and shared plates are under-represented. The hardest case for an estimate made from a photograph is a plate somebody else assembled, where the camera cannot see under the top layer. That is precisely the case we have the fewest samples of.
- Per-plate residuals are not published in the first run. Only the summary error is. Until the residuals are out, nobody outside this desk can re-derive our statistic, which means the run is checkable in principle and not in practice.
- The rig has not been round-robined. No second lab has measured our reference plates. That is the standard we apply to other people's accuracy claims and we do not yet meet it ourselves.
This entry stays open until the inter-operator check and the residuals are published, at which point it gets a closing entry rather than a deletion.
Telling us we are wrong
The fastest route in
corrections@referenceplate.com. Send the figure you think is wrong, the page it is on, and what you have that says otherwise. If you are the company being measured, send us your own measurement and how you made it.
We re-run rather than argue. If the re-run agrees with you, the page changes and an entry appears here with a date on it. If it does not, we will tell you that too, and show the numbers.