Published finding
Does the ridge decide the answer?
Every geometry we publish is g = (Σ + λ·I)⁻¹, with λ = 0.01. That constant is ours, not the data’s. A fair question — and one put to us directly — is whether our findings are properties of the data or of the constant. Answered below, on all nine corpora, including where the answer went against us.
The criticism, in the reader’s own words
Six domain readers were handed three draft pages and no other context. Four of them, from three different directions, arrived at the same objection:
“Low-variance and near-collinear columns blow up the precision diagonal… That is not nine findings about nine datasets. That is one property of inverting an unstandardized covariance matrix, reported nine times.”
It is the sharpest technical objection anyone has made to this instrument. It is also testable, and answering it first required publishing six geometries that had never been served — six of nine could not be checked by anyone outside this building. blind domain reads · 2026-08-20
The test
For each card: take its published clean state, replay its own declared red arm — the exact perturbation recorded in its localization file — and compute |Δg| at nine values of λ from 1e−6 to 1. Then ask two separate questions, because they have different answers.
Where does it move? If the strongest-moving coordinate is a property of the ridge, it should wander as λ changes. How far does it move? If the magnitude is a property of the ridge, it should track 1/λ.
The result
| card | shape | strongest movement | stable across | magnitude | ≈ 1/λ? |
|---|---|---|---|---|---|
| reading /sweep.json… | |||||
What holds
The locations, on every card. Each card’s strongest-moving coordinate is stable across at least four and a half orders of magnitude of λ, spanning the shipped value. On the question of where the movement is, the objection is answered — nine times out of nine.
What does not, and is withdrawn
Two magnitudes are not effect sizes, and we withdraw them as such. — both sit at ≈ 1/λ. They are exactly the two cards carrying a constant column, whose precision diagonal is pinned at 1/λ by construction — so “how far it moved” is bounded by, and effectively equal to, the reciprocal of a constant we chose. At λ = 1e−6 the same experiment “measures” 999,980.
The honest claim for those two is that the withheld records make a previously-constant column vary, and the instrument locates that column reliably. That is a detection with a stated mechanism. It is not a measured size, and it should not have been quoted as one. /the-record/ · 2026-08-21
What the sweep turned up while answering something else
Three of our receipts publish one measurement. agent-trace, keep-out and take-out carry the same clean-state digest, the same perturbed-state digest, the same strongest-moving coordinate and the same magnitude — to every published decimal. We replayed both of the other declarations to check, and they reproduce.
The three ask genuinely different questions of that evidence: whether an agent’s actions leaked into its record, whether a readable record shaped what a system learned, whether an erased record actually left. One experiment can bear on three questions. But three kits and three receipts, none of which said it shared its computation with the other two, presented one result as three. Each receipt now names the other two, inside its own fingerprint. /the-record/ · 2026-08-21
Not swept, and why
Every digest we publish is either in the table above or listed here with its reason. There are no silent gaps; the build refuses one.
Check it yourself
All nine geometries are served, so this does not require our bench. The table above renders from sweep.json — no number on this page is typed. Each card’s vectors and its published digest are downloadable from its case, and the sweep itself is about thirty lines of numpy.
This page exists because habitat publishes what it finds, including about itself. The test was run because four readers said the instrument might be measuring its own regularizer. On seven cards it is not. On two, the magnitude was — and saying so is the whole of the rule.