Metaphor
A word like seed, drift, threshold, or surface brings a shape with it. It suggests movement, pressure, boundary, or growth before any formal definition appears.
Concept surface / June 2026
Bathysemantics
Bathysemantics studies how wording changes what a language model notices, assumes, and does.
It focuses on the meaning beneath the literal request: metaphors, analogies, names, and repeated terms that quietly steer interpretation.
01 / Observation
Small wording changes can change the work.
Ask for a summary and a model compresses. Ask for the structure underneath and it reconstructs. The literal task is close. The mode of work is different.
Bathysemantics starts at that difference: the gap between what words say on the surface and what kind of thinking they invite.
02 / Definition
A way to study meaning below instruction.
Bathysemantics looks at the meanings that sit beneath a prompt, label, role, or interface term: metaphor, analogy, naming, resonance, and drift.
The central question is simple: what does this wording make easier to see, assume, continue, or ignore?
03 / Example
Two prompts can ask for nearby things and still steer differently.
Compression
Summarize this.
This points toward reduction: keep the main points, remove excess, make the material shorter.
Reconstruction
Find the structure underneath this.
This points toward depth: infer arrangement, dependency, pressure, and hidden order.
The explicit request is similar enough to compare. The semantic direction is different enough to matter.
04 / What to look for
The clues are often ordinary words.
Analogy
An analogy moves relations from one field into another. It can make a system easier to understand, but it can also smuggle in the wrong expectations.
Naming
A good term reduces explanation cost. A weak term stays flat, or drifts until different people and models no longer use it the same way.
05 / Uses
Useful where language becomes part of the system.
The work applies to prompts, evaluator wording, role names, taxonomies, memory labels, governance terms, product names, and internal concepts.
The goal is not richer prose. The goal is clearer control over the meanings a system is already using.