Beyond Short Names: Emoji Keyword Retrieval across Models and Bounded Causal Tests

Abstract

Revised independent preprint, version 1.0.0. This preprint was revised in response to reviewer feedback from the NeurIPS 2026 workshop Interpretability as a Science (InterpScience). It is not an accepted workshop paper or a conference proceedings paper. We distinguish information readable from hidden states from its causal contribution to a specified computation. Using emoji as a controlled text-only case, we prospectively separate cross-model keyword retrieval, component-local mediation, and attention-path contributions after a readout reset. We compare raw, official-name, word-disjoint CLDR-keyword, and codepoint views of 55 emoji and 55 non-emoji Unicode symbols. Across nine prespecified model-tokenizer configurations, emoji keyword-minus-codepoint retrieval exceeds the registered 0.02 floor in unrestricted and exact four-view token-count-matched analyses. Keyword-mapping randomization passes the four-model partial-conjunction primary (p = 0.0027). Symbols show broad official-name retrieval and smaller, heterogeneous keyword effects. In Llama 3.2 3B, rank-8 subspaces learned from 128 source emoji satisfy no primary late-component joint necessity-and-sufficiency criterion on 192 held-out emoji. A separately specified optimized-subspace follow-up on 384 further emoji returns uncalibratable at both primary residual sites. Its full-state controls reveal a depth-dependent asymmetry, motivating an attention-path test. On 195 additional, prospectively specified items, glyph-position message disruption and restoration after readout-state replacement shift keyword cosine margins beyond equal-count suffix controls; all five simultaneous one-sided 95% lower bounds are positive. Absolute glyph-message disruption and restoration have zero mean top-1 effects. The results establish cross-model lexical readability and a conditional attention-message contribution. The tested rank-8 account remains unsupported. A subsequent exploratory response-scoring extension on 24 authored scenarios exposes wording and polarity failures in the proposed measures. A prospective text-only qualification on 32 new facts then yields 88/96 correct reserved direct/minimal answers, four errors, and four ties, failing its all-cases-correct criterion. Separate label mappings also fail their criterion. Contextual interpretation and output-linked causality remain unresolved. Reproducibility scope: the accompanying package contains the current manuscript, selected public inputs and aggregate evidence, a preserved historical compact archive, and all 4,640 score records for the explicit-report qualification with two saved-score validation implementations. T19-L and T21 support specified statistical replay; T19-C/T20-C provide summaries rather than complete item-level inference replay. Model weights, raw historical activation arrays, private individual author responses, and the original private repository history are excluded. Saved-score verification is not an independent model experiment or human validation. Licenses: the paper, its original figures, and the newly released project-authored data are CC BY 4.0. Project code and ordinary project documentation retain Apache 2.0. Third-party and historical materials retain their original licenses and notices; see the package's file-level license register. No checkpoint or tokenizer assets are redistributed.