How The Vault Report Researches

Every figure on this site is read from the file the institution itself published, then checked against that original before it ships.

The Vault Report is the precious-metals and commodity-inventory desk published by ByShovel. We pull from the institutions that publish the data, never from sites that repackage it, so the number on our page is the number in the official report and the original is the citation.

The Vault Report exists to break commodities stories that are new and can be checked. Not a restatement of a figure the exchange already publishes. A finding nobody else has, with the evidence on the page.

The raw material is our own capture: firm-level COMEX deliveries, open interest the exchange does not archive, warehouse stocks read from the source file each day and reconciled against the day before. We cross that with a ledger of the claims circulating about metals, so we know what has already been said and where it is wrong.

The site is also the public demonstration of ByShovel's engineering. If the pipeline cannot be trusted, neither can the story, which is why the method is on this page.

There is no staff, no editor and no review board. When a page says “we”, it means one person and the systems he built.

When we say a level is the lowest since a given year, or that a spread has crossed a threshold, that claim is anchored to a dated reading you can verify yourself on COMEX, LME, oil or the SPR.

When a piece was checked, it says how. The usual shape is a full scan of our own archive, then a fresh pull of the live source to confirm the two agree. A number that fails that check does not ship.

When something is not knowable from the data, the piece says exactly what is unknown and carries on with what can be said. A hedge is not a reason to stall.

How a figure gets published

Ingest the raw report from the source on its own schedule.

Reconcile it against the prior reading, so a bad parse or a stale feed is caught before it ships.

Publish the live number with the date it covers and a link back to the primary source.

A claim is only as long as the archive behind it. “Record” on this site always names its window. A capture that began this year cannot refute a twenty-year impression, and we say so rather than pretend otherwise.

Other series sit between those bounds. Wherever we use record language, the span comes with it.

When we get something wrong, we fix it on the page where it was published. The correction is dated and says what was wrong. The original wording stays below it, so you can see what changed. We do not quietly rewrite.

Forward-looking statements ship with what would prove them wrong and a date on which we come back. The revisit goes on the same page, dated, and says plainly whether we were right.

We keep our own error log, including the times a fix overshot in the other direction.

To report an error, write to [email protected] with the page address and what you think is wrong. A source helps.

AI systems run the capture. They fetch each source file on its schedule, parse it, and reconcile it against the prior reading. A figure that fails that check does not publish.

They also draft. Before a draft reaches a person, a verifier checks it against the doctrine: every figure must trace to a row in our archive, record language must carry its window, and a claim about a named actor must rest on a finding about that actor. A draft that fails the check does not go forward.

What AI does not do is decide. A research article or claim-check publishes only after a person has read it and signed off. The house position comes from the doctrine, which a person wrote. And corrections are made by a person, in place, as above.

The live data pages and the alert emails are the deliberate exception. They update on their own once a figure clears the checks, because the point of them is to move the day the source does.

One address for everything: errors, story tips, data questions and licensing. The founder reads it.

Back to the dashboard