Coverage

Kestrel does not choose which permits it is measured on. A forecaster who picks their own benchmark can always look good, so the denominator on this page has to be somebody else's list.

The denominator was withdrawn on Sep 4, 2026. Until then this page carried a count of the permits named as pending in mining companies' filings to the American securities regulator. A hand-check of 60 of those rows, read one at a time against the filings themselves, found fewer than a third correct as stored: some were labelled wrong, some were counted twice, and some were not permits at all. A replacement is being built, a census of development-stage mining projects in Canada, the United States and Australia. Until it exists this page publishes no coverage percentage, because a number we have just found to be wrong is worse than no number.

What we cover, register by register

Each row is one public register and Stage is where it sits in the pipeline described below. Within each group the registers most often named by companies come first.

There is no count beside them. The counts this page used to print came from the same withdrawn extraction as the headline, so they carry the same error rate and are not published.

The distribution is lopsided and worth saying plainly: the permit types we publish forecasts for account for a small share of what companies name, and the registers carrying the most pending permits are ones we have not assessed. Closing that gap is the work.

It got narrower before it got wider. On Aug 13, 2026 five of the six permit types then being published were withdrawn from the record, because they had been promoted through a review the principal judged inadequate. Those rows are below, with their reasons on the Changelog. One permit type is published today.

The pipeline, and the two public stages

A permit type reaches this site in four steps. The first two happen out of sight; the last two are public, and are the badges shown here and on the Track record tab.

Not started. We know the register exists and companies are waiting on it. Nothing is built.

In development. The register is captured and the permits are in our database, but there is no published forecast. Model work happens privately, and no version of it is shown here: a model that never went public cannot be quietly resurrected as evidence later.

Testing in public. Forecasts are published, dated and scored, but the record is still too short to conclude anything from. Everything published at this stage is scored on exactly the same terms as everything else, and none of it is withdrawn for scoring badly.

Standing record. The record is long enough to stand behind. The rule, fixed in advance: at least 30 decided permits, and the range of true coverage consistent with that record still contains the 80% we claim. Promotion is announced on the Track record tab when it happens.

Nothing skips a stage, and nothing moves backwards quietly. A family whose coverage deteriorates after promotion stays on the page with its record intact.

Why the denominator is not ours, and what happened to it

The rule stands whatever the number is: the list we are measured against must be drawn up by somebody other than us, and permits we do not forecast must stay in it and hold the figure down. That is the whole point of publishing a coverage figure at all.

The first attempt at such a list was every permit named as pending in mining companies' filings to the American securities regulator. It was somebody else's list, which was right, and it was wrong in three other ways. Its unit was mentions rather than permits, so one permit named in three filings counted three times. More than half of it was permits nobody had applied for yet, which no timing model can forecast. And American filings describe the American-listed world rather than the Canadian development projects a royalty buyer actually holds.

Then the hand-check settled it. Sixty rows were read against the filings they came from: seventeen were right as stored, fourteen carried the wrong label, sixteen were not permits at all, and eleven were a permit already counted. Neither the extraction's own confidence score nor its verbatim-quote check told the good rows from the bad. On that rate the real figure is nearer two thousand than 4,337, before counting the same permit twice across filings.

So the count came off this page on Sep 4, 2026 rather than being adjusted, and the replacement is a census of development-stage mining projects in Canada, the United States and Australia, which is a list of projects rather than a list of sentences about projects. When it exists the figure returns, and it will be published every week whichever way it moves.