Dimension registry

A score that travels without its status eventually gets quoted as a result. The registry is the artifact that stops that: one machine-readable entry per dimension, carrying what the dimension measures, how it is scored, and what evidence stands behind it. A consumer that cannot read a status can refuse to render the number.

The registry is a contract, not documentation

Prose describing a dimension can drift from the code that computes it. A registry entry is read by the thing that renders the score, so a wrong entry breaks a surface rather than merely misleading a reader.

What one entry carries

Field	Holds	Why it is required
Key	The immutable identifier, for example $\xtask$	Joins a score back to its definition across versions and datasets
Name	The human label	What a surface prints; may change freely, unlike the key
Question	The single question the dimension answers	A dimension whose question cannot be stated in one sentence is more than one dimension
Value space	The permitted values, including $\abst$	Lets a consumer distinguish a real zero from an absent measurement
Method	Model-judged, rule-scored, or a fold of both	Determines what calibration would even mean for this dimension
Evidence status	What the evidence is, and what it was measured on: fitted, preview, or none	The claim about validity, defined on the scoring standard
Promotion status	Whether the dimension is in force: live or provisional	A separate axis from the evidence. A dimension can carry strong evidence and not be promoted, and a promoted one can rest on none
Lifecycle status	Current or superseded	Which definition is in force
Calibration reference	A pointer to the record, or an explicit null	An absent record is a fact worth publishing, not a blank
Supersedes	The prior key and version, where one exists	Keeps an old figure traceable to the definition that produced it

What a calibration record itself must contain is specified on calibration, not here. The registry points at that record; it does not restate it.

Rules that make the registry trustworthy

A key is immutable and never reused. Renaming a dimension in place silently rewrites the meaning of every score already published under that key. A changed meaning is a new key that supersedes the old one.

A superseded dimension stays in the registry. Deleting it orphans every figure that cited it. Retention is what lets an old result stay readable without being mistaken for a current one.

Every dimension carries both statuses, always. The common failure is a dimension that is current and provisional at once, published with only the lifecycle half visible, so a live definition reads as a validated measurement. Both axes are independent and both are required.

A null calibration reference is written, not omitted. The same discipline as abstention: the absence of evidence is recorded as an absence, never as a missing field a reader fills in optimistically.

Promotion

Promotion runs one direction, and only measurement moves it. Shipping a stronger model does not promote anything, because the missing thing is evidence rather than capability. The conditions a dimension has to clear live on the validation protocol, which is their one home. This page does not restate them, because a second copy of a rule is how the weaker version gets quoted.

The registry is where that rule becomes enforceable rather than aspirational: promotion is an edit to a status field with a required calibration reference, so a dimension cannot be promoted without naming the evidence that promoted it.

Named gaps

The registry describes dimensions, not the tasks they are computed over. A biased task distribution produces well-formed entries with sound statuses and misleading scores, and nothing in the registry can see it.

Registry entries are also per-dimension, not per-surface. Two consumers can read the same entry and present it differently, so a status shown correctly in one place can still be dropped in another. The registry makes the status available; it cannot make a surface honor it.

The lifecycle rule above is in force and unmet, which is recorded rather than softened.

Conformance gap: no registry carries a lifecycle field

No dimension registration in either scoring system has a lifecycle field, so the requirement that every dimension carry both statuses is unmet on every entry rather than on some of them. One system tracks a lifecycle for guides and for judgements, never for a dimension.

Nothing has been superseded yet, which is why this has cost nothing so far and why it is worth closing now. The cost of adding the field is lowest before the first supersession and highest after: a key retired without a lifecycle record orphans every figure that cited it, and the record cannot be reconstructed once the definition is gone.

The Supersedes field is missing for the same reason and is not separately marked: no entry in either system carries it, because neither has the per-key supersession this page specifies. One versions the whole registry instead, minting a new registry version when any meaning changes, which keeps old figures readable but cannot say which single dimension moved or what it became. That is a different mechanism rather than a partial one, so closing this means choosing between them, not filling in a blank.