The Head of the Swarm: Field Tests of Tagged Zeros on Earthquake Catalogs and a Language Model

Our previous paper split the unresolved verdict U of a ternary test into three tagged zeros: a data zero, decided bymore of the same data; a model zero, decided only by changing the question or the instrument; and a principle zero,decided by nothing within the declared hypothesis families. Here we take the tags into the field. We tagged the claimsthat had not stood up across the WCM programme (28 claims: 19 clean F, and 9 U) and then settled four of the Uverdicts with sealed tests on earthquake catalogs (USGS, JMA, NIED F-net) and on hidden states of Llama-3.2-3B.Three findings stand. First, for count zeros the depth tag was right but the direction read from the point estimate waswrong in all three cases checked (n = 3). Second, data zeros hid model zeros. A residual "orientation-specifictemporal clustering" of focal mechanisms (z = +1.48) vanished when only cluster-initial events were kept: z = -1.06,below the 5th percentile (+0.63) of random thinning to the same size. The residual was the similarity of events insideone cluster. Third, a novelty reader (dissimilarity to the preceding K events) turned out to be a cluster-head detector.M7+ mainshocks stood out (58.1th percentile, sealed), and so did cluster-initial M5.5-6.4 events on unseen data(57.7); a size-specific excess was not seen (D = -0.84, U). Inside the language model, clause-initial tokens carry thesame relational mark: it is built in middle layers (peak at layer 12) and absent at the embedding layers, and the wordafter a clause head shows the opposite sign. The same mark is lost in the top four principal components, whichexplains an earlier null result of a 24-cell ternary reader (a model zero). We also report two flaws we found in our owninstruments: the reader's sign layer depended on arbitrary axis orientations, and once made orientation-invariant itseparated any two lexically different token groups. Recovering an original instrument from its source reversed one ofour own interpretations. Everything that rests on post hoc analysis is marked as such. (1) Text: CC-BY-4.0 | Code: GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later) (2) Data and code. Seal documents with their SHA-256 files, result files, development logs, the sentences and theextraction script, and all analysis scripts (GNU Affero General Public License v3.0 or later) are provided asBridge_Paper_The_Head_of_the_Swarm_scripts_v1.0.0.zip. Catalogs and hidden states are not redistributed; theyare available from the sources above, and the hidden states can be regenerated with the included script. Llama 3.2is licensed under the Llama 3.2 Community License; built with Llama. (3) AI use. Designs, code, analyses and text were produced by Kurado (Claude Opus 5.5, Anthropic) working with M. Watabe, who chose every direction, supplied the data and ran the language-model extractions.