Surfaces Do Not Reach Addresses

A knowledge store that computes its addresses from content promises that entities it never saw during training remain reachable. We test that promise on real data — 94,690 WikiData people, four relations — and find it splits in half. Addressing survives. Keys computed from canonical names reach never-supervised entities at 0.999560, indistinguishable from a matched synthetic control (+0.104 points against a 1.0-point gate, standard error 0.016). It holds across a 47.3× increase in corpus size, over which recall rises while the median identification margin falls 45.6%, and across an 8× sweep of latent width: effective rank sits near 22.575 at every width, so utilisation is not a constant and address space cannot be bought by widening. Resolution does not. The missing invariance is not in the encoder and cannot be trained in — a raw alias surface reaches its entity 0.0475 of the time, and adding alias pairs as contrastive positives failed to fit 0.9516 on its own training data. It cannot be prompted in either: a 1.5B instruct model, under one prompt, canonicalises 0.0155 of variant surfaces, and instructing it to resolve multiplies the undetectable-error rate 7.2×. Descriptive access is closed outright — 0.0060 correct against 36.4 : 1 confidently wrong, on an optimistic read drawn from the identifiable end. What works is lookup: indexing the mapping restores the end-to-end composite to 0.9540, for the surfaces the index contains. The write path answers this by declining. Misplacement — permanent, confident, invisible to any read-time threshold — is held to an observed rate of 0.0021 on all four relations, every 95% upper bound below 0.0056; unresolvable subjects are quarantined rather than placed. That path missed its own registered gate, reported unadjusted. The methodology is a co-equal contribution: two of six inherited species of instrument failure produced fresh instances here, and one species is newly named — the orphaned derived figure, a value that is correct while its stated relation to its operands is not. (No capacity claim, no corpus-level acquisition claim, and no claim about other model classes — §9.) This deposit names internal commit 26636247ca7b3cbbe9830d7d4a05bd277dab6a1c as the state of record (not publicly obtainable). The enclosed source archive is the curated export supporting this paper's claims; its MANIFEST.json maps every archived file to its content identity at that commit. Companion repository: https://github.com/Vexillon-ai/Athena. Enclosed code is MIT-licensed; this deposit's documents are CC-BY-4.0.