Words Are Not Containers
The common picture treats a word as a box with a meaning inside. This paper tests a different idea: that meaning is relational and lives in how a word is used. We look inside open transformer models to see whether their internal states track that difference.
- Cross-model replication. The core signals, including model "fingerprints" and an idiom-vs-literal separation, replicate across three independent 7B model families.
- Semiotic accuracy. Internal representations separate figurative from literal use well above chance in every model tested, and the effect is statistically strong in all three.
- Human baseline. A blind human annotation study is under way so the model results can be compared against human judgment on the same items.
- Open questions. Some follow-on experiments have only run on one model so far, and the layer-by-layer analysis is being reworked before we report it.
The full experiment suite, corpora, results, and annotation kit are public. Ask us for the repository link or a copy of the current draft.