Toy models of the collapse of distinctions in a represented world, and the instrument that measures what is missing.
A represented world loses distinctions its source world holds, and the loss does not register as loss. What a system ceases to represent cannot, through that representation, generate the corrective signal needed to restore itself.
This is a selection pressure with no error signal. Distinctions that fall out of a representation generate no queries against it; absent queries produce no correction; and the metrics computed over what remains improve as the band narrows.
A recursive representational system remains open to the world only insofar as evidence not generated by the system itself retains operative standing over the system's next output.
Three reasons, and they compound.
Truncation misses most needs. Under a Zipf-shaped demand distribution, the distinctions dropped are individually rare and collectively most of the field. A system serving the head serves most queries while failing most needs, and the two counts diverge without any measure registering it.
Noticing requires already knowing. A reader who does not hold the lost distinction cannot detect its absence. The people positioned to notice are precisely those who least need the system, and their non-use removes the only signal that would report the gap.
Defection is hysteretic and scoped. Users who do notice leave for that purpose and return for others, so the aggregate usage signal barely moves. The departure is real and does not appear.
A general metric cannot see this, because every general metric is computed over the distribution that has already contracted. So the measure is against a reference inventory fixed before any query runs.
For each item: the distinctions the source world actually holds, written down first. Then the address is queried and the returned distinctions counted against that list. DS — distinctions surviving, over distinctions held.
The panel holds thirteen items across ordinary polysemy, technical terms, proper names and literary fragments. Bank has three senses in its inventory; a naked address returned one. DS = 0.33.
The reference inventory is the whole instrument. Quality raters judge the answer shown, not the inventory unshown — which is why a composition can score full marks while carrying one sense of three. Rating output quality is not measuring coverage, and nothing in an optimization loop measures absence.
The panel's first battery reported a recovery rate of 0.5 — distinctions lost at an address and reached again later. That number was inflated by counting the wrong thing.
Of three recoveries, only one occurred at the same address form. One item lost a distinction at a naked address and recovered it at an adjacent one; another recovered at an entity-tied address and at a typo, while the naked address returned the flattened term three times across three weeks.
That is not temporal recovery. It is a graded membrane: the entity-tied address carries what the naked address drops. At the address ordinary users type, observed recovery is one of three, and the one case with repeated naked sampling shows none.
Contraction is reward-positive and recovery is reward-negative, which makes this a ratchet rather than a dial.
A reward computed over the current distribution cannot see what left it. Specialization within the remaining band raises every score. Restoring tail distinctions spends capacity on material the evaluation set rarely contains, so it registers as regression on exactly the instruments used to decide.
And it compounds from the demand side: people stop asking what the layer cannot answer, so the query distribution contracts, so the evaluation set contracts, so the target narrows again.
DS stable or rising across the panel over a sustained interval, at the same addresses under the same conditions, would disconfirm the contraction claim for the measured items.
A general metric shown to detect distinction loss without a pre-fixed reference inventory would make the instrument unnecessary.
And a demonstrated recovery mechanism operating at naked addresses — distinctions restored where they were lost, not merely reachable at adjacent forms — would break the ratchet claim.
The line may not be defended by any single flattened term. Thirteen items on one instrument is a beginning, and the panel is one day older than its first result.