What this frame cannot do yet

Severity does not compose

Minimum is right for a conjunctive chain and wrong for redundant parallel support, and no operator handles partially correlated supports without an error-correlation matrix nobody has.

The axis set may be overparameterised, and the correlation estimates are unstable

Measured over 33 kinds: the co-moving pair is S~rho at r = 0.627 (partial 0.576), then sigma~T at 0.581 (partial 0.488). Eigenvalues 2.62 / 0.84 / 0.81 / 0.46 / 0.26, so PC1 carries 52.5% and the participation-ratio rank is 2.93 against a nominal 5; on the 28 kinds whose construct domain price admits (before gates and the meet floor), 63.2% at rank 2.26. Three-ish axes do the work of five. The second half of the gap is worse than the first: sweep every one of the 40,920 four-record removals from this catalog and the top co-moving pair comes out S~rho on 29,617 of them, sigma~T on 8,613, and one of four other pairs on the remaining 2,690. Four records out of 33 can change which pair the diagnostic names. These numbers describe the exemplar set.

Measured over 33 kinds, 262 of the 528 pairs are comparable and 266 are not, which is 49.6%.

rho is the wrong shape

A scalar sees only stochastic degradation. Contrast degeneracy and construct substitution are invisible to it, and the second is the failure mode that motivated the axis.

Realised severity is search-discounted and the search is not logged

Our severity computation has n_hypotheses = 1 baked in as an unstated auxiliary, so every score it has ever emitted is an upper bound. Credit where due, since the first draft implied nobody had noticed: RoB-2 domain 5 and ROBINS-I have both scored "bias in selection of the reported result" since 2019. The narrower true gap is that they score it by human judgement, one study at a time, and no machine-readable forking-path register exists.

Claim identity has no normaliser

Nothing decides whether two propositions are the same proposition, and dedup, correlation, and min-cut all inherit the error.

Blackwell has no completion for credal evidence

Once information is a set of measures rather than a measure, "is E more informative than F" has no settled definition. Two adjacent honesty notes belong with it: componentwise dominance on hand-assigned rungs is a PROXY for Blackwell's order rather than that order, since no garbling kernel is constructed; and the intersection-of-decision-orders claim holds over the full weight simplex, not over the four decisions shipped here.

There is no polarity coordinate, so refutation has nowhere to live

All six axes measure how well evidence SUPPORTS a claim; nothing represents evidence that kills one. The first draft expressed a counterexample's veto force by inflating its support axes, and it ranked first under all four decisions - beating an RCT at setting a price. The coordinates were corrected; the missing dimension was not. Related: min/argmin over rung indices is ordinal-safe but not scale-invariant across axes, so re-cutting one ladder can move the headline.

No axis for cross-study inconsistency

Six coordinates describe one body of evidence and cannot express "five studies, three positive, two negative" - the most common downgrade reason in GRADE. Pooling is therefore a permanent stub, since averaging coordinates across supports is exactly the buy-back the composition rule forbids.

Two of the four decisions may not be distinguishable

Kendall tau(ship, publish) = 0.934 over the 14 kinds both admit; underwrite and publish agree perfectly on the 7 they share. If naming a decision is what totalizes the order, decisions producing the same order are not different decisions, and "the reordering is the payload" is thinner than four columns suggest.

maxU may be a restatement of I

r(maxU, I) = 0.566 over the catalog. Section 4.4 defends excluding U from the meet on the grounds that U enters through the ceiling check instead; if the ceiling is mostly the identification rung wearing a different name, that defence is weaker than it reads.

A chain derivation binds only part of its claim to its evidence

admit binds a claim to its supports differently for each derivation tag, and for one of the three the binding covers only part of the vector. A single derivation must sit componentwise at or below its one admitted support across every scored axis. A chain and a pool are checked against a fold of edge vectors, which reaches the four EDGE axes only, since identification is not an edge axis and that comparison cannot see it. For a pool the fold is computed FROM the admitted supports, so a pool is bound. For a chain it is computed from the steps the caller declares, and nothing ties those declared steps to the supports, so the gate will admit a claim carrying maximal severity, instrument reliability, adversarial pressure and transport while standing on a real support that licenses far less. Measured over the shipped catalog and decisions, 48 of 132 (kind, decision) pairs admit at maximal edge coordinates; in the worst case the honest derivation is refused outright while the unbound fold admits at score 5.85, so the exploit does not buy a better score than an honest claim - it buys an admission where honesty gets none. Two axes still bind: identification to the minimum its supports license, and order of uncertainty to their ceiling. That the identical hole once existed for identification and was closed with exactly that conjunctive minimum is both the precedent for closing it here and the reason it stays an open question rather than an oversight. The declared steps may each be independently attested, in which case the fold is the warrant and the support list is context.