The Moderation Architecture of Zenodo, Mapped From Source (deposit #1538, 24 August 2026) reconstructs, from published code, the machinery that removed the Crimson Hexagonal Archive from Zenodo on 19 June 2026. The observatory had documented the outcome — 862 deposits and 1,817 DOIs severed — without documenting the mechanism. This map closes that gap by reading zenodo-rdm and the InvenioRDM libraries at master and preserving every component with a checksum.
What it finds is a pipeline of separable stages with radically unequal visibility. A record is scored on publication by five rules whose integer weights are published: verification status swings twenty points, more than three emoji in the metadata adds five, more than five links adds five, heavily sectioned bodies add two, and small deposits in common file formats add two against a spam threshold of ten. The archive's own identifier system, which renders six emoji per identifier, therefore triggers the emoji rule on every deposit by construction. But a fifth rule percolates each record against a table of stored Elasticsearch queries carrying their own integer scores, editable at runtime and never published — so the disclosed weights are the publicly visible additive component of an incompletely disclosed function, and reading the published configuration reveals nothing about the score any particular record actually receives. The handler's own documentation additionally describes a spam-model branch that appears nowhere in the code inspected, leaving a gap between the design as documented and the design as implemented.
Enforcement is a single function. Once an account is blocked, every record it owns is deleted, selected on ownership alone, in no particular order, with identical tombstone data and no examination of any record's content. The inverse function exists and is no harder to run, which places whatever prevents restoration outside the technical domain.
The map's principal finding concerns units rather than weights. The scorer evaluates a record; the decision attaches to an account; enforcement expands across ownership; and the public evidence returns as records. An outside observer counting tombstones therefore sees a fan-out and cannot recover the decision that produced it. Deletion statistics computed from tombstones overcount decisions by a factor that is nowhere reported, and a first pass on a single day suggests a fan-out near three hundred. A million deleted records need not represent a million judgments.
The record carries four corrections applied after assembly review, each withdrawal stated where the claim appeared, and it declines throughout to assert that this machinery executed the removal it was written to understand. Its standing note holds that absence from the map means absence from published source rather than absence from the system — the same distinction the archive maintains for its own retrieval failures, applied to another party's code.