GRAYSTAK

How we work, and how to check us

Version 1.0, September 2026. A finding is only worth relying on if you can see what was examined, what was left out, how confident we are, and whether our earlier calls held.

What every piece carries

Most of this already governed our published work. Three requirements are newer: a stated question before the evidence, a confidence drawn from a fixed scale, and a public registry entry for anything forward-looking.

A question
What is being asked, in one sentence, before any evidence is gathered.
A method note
What was examined, what was excluded, and what was transformed. Sources named.
Findings
What the record shows, each tied to a document, a proceeding, or an observation.
An assessment
What follows from the findings, with a stated confidence and the reasoning.
What would show it wrong
The conditions under which the assessment fails, dated where a date exists.
Implications
Who this reaches, on what horizon, and what decision it bears on.
A registry entry
Logged with an identifier, a date, and a confidence. Updated when it resolves.

The separation between findings and assessment is the load-bearing part. A finding reports the record. An assessment reasons from it. Conflating the two is how a reader ends up unable to tell which parts of a document they can check.

The confidence scale

Documented
Rests on a primary document we hold and cite. Overturned only by that document being wrong.
Inferred
No single document states it. The mechanism is established and every input is documented. Overturned by a named document appearing.
Contested
A supported reading exists and so does a competing one. Both are stated. Resolved by a ruling, a filing, or an enacted text.
Unresolved
We cannot tell, and say so. Carried forward as a gap until it closes.

We do not publish probabilities. A percentage implies a base rate that does not exist for most of what we cover — there is no reference class for a Section 232 report that is two hundred days late. Stating the evidentiary basis, and the specific thing that would overturn it, tells you more than a number would, and it can be audited afterward.

The method note

Every piece carries a short note stating what was analyzed and what was not. At minimum it names the sources examined with dates of access; what was excluded and why; any transformation applied to figures, including how a rate, a spread, or a day count was computed; claims removed during verification with the reason each was removed; and requests for comment with the date sent, the deadline given, and whether a response arrived.

The fourth of those is the one most likely to be skipped and the one that does the most work. A published note saying which favorable claims were dropped, and why, is the cheapest available demonstration that the rest was checked.

The registry

Every forward-looking assessment is logged with an identifier, the date it was made, its confidence, and the date or condition on which it resolves. When it resolves we record the outcome, whether or not the assessment held. Assessments that fail stay in the registry at the same prominence as those that hold.

This is not a separate system. It is the forward call registry that already governs the index, extended to research. One registry, one pre-registration rule, two classes of entry.

Written first
An entry is written before the resolving date, never after.
Failures stay
An assessment that fails is marked failed and remains visible. Nothing is removed.
Read as a whole
One call resolving as predicted is not evidence of method. The registry is read together.
Partial resolutions
Where a call resolves partly, the entry says which part and why.
Nothing counts
A date that passes with no document is a result, logged with the same weight as one that produces something.

When we are wrong

Corrections are dated, visible on the piece, and describe what changed rather than only that something changed. A correction that changes an assessment triggers a registry update the same day. Where a figure is revised, both the prior and revised values are shown, including when the revision favors our own argument. We do not silently amend.

What this standard does not cover

It sets how work is documented, not what is worth covering — subject selection is an editorial judgment and stays with the editor. It does not set the validation protocol for the scored signal product, which has its own calibration, freeze, and null-selection rules; the two share the registry and the pre-registration discipline, not the methodology. And it does not make a finding correct. It makes a finding checkable, which is a different and more modest claim.