The Nature of Knowledge: six axes, and what each licenses alone
Six axes type a claim and the evidence behind it, and none of the eight pages under this section, including this one, licenses a claim on its own. Two axes describe the claim’s own shape. Four describe the edge between the claim and whatever produced it. This page is the complete short read: the six axes, why they compose by minimum and not by average, the gate that runs before any of that, and why the section beneath it is split the way it is.
Two axes about the claim’s own shape
U and I are the two axes most often collapsed into one, and they ask different questions. U types the shape of what a report asserts, a point estimate, an interval, a distribution, a partially identified set, running from maximally committed to explicitly bounded. I types whether the quantity the claim is about is recoverable from evidence at all, independent of how carefully the claim is stated. Neither buys the other: climbing U cannot rescue a confounded estimand, and a well-identified quantity can still be reported as a bare point with the error bar deleted. U alone sits outside the rule, covered below, that folds the other five axes into one number.
Four axes on the edge
The other four, severity, instrument reliability, adversarial pressure, and transport, are not properties of the claim or of the evidence taken alone. Each types the edge between them: how hard the producing check tried to break the claim, whether the instrument that read it can be trusted, whether whoever checked it was independent of whoever produced it, and whether the evidence’s own region reaches where the decision lives. That is what lets the frame compose at all: a property of an edge can be scored without settling what the claim itself is worth, then combined with the other edge properties into a single bound.
Each axis’s one-line question, and its orthogonality witness pulled live from the evidence catalog: a real record for five axes, and a hollow ring for U, since the package states plainly that U types reports, not evidence kinds, and no catalog record can witness a property evidence does not have.
Pricing figures throughout are synthetic and chosen to make the type of the claim legible; none of them are measurements.
A worked example clarifies what the figure’s U row is doing. U0 in the package’s own ladder is the point estimate "Elasticity is -1.8.", an error bar deleted rather than a small one. Reporting that same number to six decimals would not move it up a rung; U measures the shape of the commitment, not the precision of its display, and no evidence kind in the catalog can serve as U’s witness because U is a property of how a report is shaped, not of the evidence behind it. The ring around U’s marker above is drawn hollow on purpose, not left off.
Trust meets, it does not average
Composition folds the five scored axes, I, S, rho, sigma, T, by taking their minimum, never their average. Averaging would let a claim buy back exactly what a single weak axis is supposed to block: a claim strong on reliability and weak on severity should not out-rank one with the two numbers reversed just because an average cannot tell them apart. The composite is the greatest sound abstraction, the tightest bound that still never over-claims, and the full argmin set, not a single number, is the honest headline.
The construct-domain gate comes first
Before any of those five axes gets scored, a construct-domain gate asks whether the evidence’s domain, empirical, formal, or synthetic, matches what the decision admits. A domain mismatch is not a discount to apply after scoring; it is a reason to refuse the evidence outright. A machine-checked proof carries exactly zero bits about a demand curve, whatever its reliability, and no severity or reliability score changes that.
About half the catalog is not ordered
Componentwise dominance, one kind beats another only if it scores at least as high on every axis and strictly higher on one, is the one ordering that costs nothing to justify, because it never names a decision. Measured over the 33-kind catalog, 262 of 528 pairs are comparable and 266 are not, which is 49.6% decidable without picking a loss function. The rest need a named decision, and naming one is what the 4 scored decisions on the catalog route each do differently.
Why eight routes, on subject-matter grounds
The section under this hub has eight children, and the split follows the subject matter, not the six axes. Five routes cover the axes between them: identification (I), severity (S), instruments (rho and sigma together, since their interaction has no home split apart), transport (T), and uncertainty (U). The other three, composition, open gaps, and the evidence catalog, are not axis pages at all; each is its own subject, not a seventh piece of the ladder.
That split is a choice about what is easiest to read, not a fact about the six axes themselves. The honest description is subject-matter grounds, not axis structure: claiming the split follows from the axes would be a claim the section’s own three non-axis children refute on their face.
The eight routes
Whether the quantity a claim is about is recoverable from evidence at all.
Licenses nothing alone.
How hard the producing check tried to break the claim, and whether it survived.
Licenses nothing alone.
Can the measurer be trusted, and was the checker independent of the producer.
Licenses nothing about the world alone.
Does the evidence's own region reach where the decision actually lives.
Licenses nothing alone.
What shape a report commits to, from a bare point up to a named residual.
Licenses nothing alone, and sits outside the meet by construction.
Why the five scored axes fold by minimum, never by average.
The rule, not a seventh axis.
11 dated gaps, each with a falsifiable prediction attached, including one that could retire two axes.
The frame's own risk register.
33 evidence kinds ordered by dominance, and what the ordering refuses to decide.
The partial order over kinds.
Rosetta Stone
Four circles, four readings of the same object. Each role reads the artifact through its own lens.
Before you rank an opportunity, rank the evidence behind its numbers. A single elasticity estimate and a preregistered field experiment can carry the same headline number and completely different admissibility - the frame is how you tell them apart before capital moves, not after.
The tell that catches every team eventually: a metric everyone repeats until it is load-bearing, with nobody able to say who checked it, on what population, or whether it transports to the decision at hand. Ask which axis is missing before the number ships to a deck.
If you are wiring an eval or a verifier into a pipeline, this is the type system underneath 'trustworthy output': six coordinates, a construct-domain gate that runs before any of them, and a composition rule that is a minimum, never an average. Build the gate first; it is cheaper than discovering the gap after a bad number ships.
The frame is explicit about what it cannot do yet, ten dated gaps, including one that could retire two of its own six axes. That is the falsifiability standard the rest of the section holds itself to, and it is worth reading before trusting any single axis page at face value.