01
Attractor-sensitive language
Why seemingly minor phrasings can shift a model from summary to synthesis, from compliance to investigation, or from surface completion to structural reasoning.
Premise
Language is not only instruction. It is also steering.
Small wording shifts can trigger qualitatively different trajectories in language models. Not just different answers, but different semantic attractors: different role assumptions, different conflict resolutions, different forms of depth or flattening.
Bathysemantics asks what beneath-the-surface structures are doing that work. Which images and tensions are already embedded in a term? Which local meanings are being stabilized? Which frames are being activated long before a system explains itself explicitly?
Where it matters
Three places where deeper meaning changes behavior.
02
Semantic infrastructure
How resonantly chosen terms reduce explanation cost by carrying image, direction, and behavioral expectation before a system is fully formalized.
03
Metaphor as control surface
Metaphor is not ornamental frosting. In pattern-based systems, it helps pre-structure interpretation, transition logic, and the field of likely continuations.
Relation to adjacent work
Not a replacement for harness engineering. A deeper-reading layer within it.
Harness engineering asks
- How do we steer, constrain, and evaluate agents?
- How do we build reliable control layers around model behavior?
- How do prompts, tools, evals, and runtime checks fit together?
Bathysemantics asks
- What meaning pressures are hiding underneath the steering language?
- Which resonances or metaphors are quietly shaping interpretation?
- Where does semantic drift begin before failure becomes obvious?
Current position
A way to read the hidden steering in language.
Bathysemantics is not a replacement for prompt engineering, context engineering, or harness engineering. It is a way to ask what the language inside those practices is already doing before anyone formalizes it as a rule.
What comes next
The concept earns itself through readings, not slogans.
Read one term deeply
Compare surface meaning with bathysemantic reading and see what becomes visible only at depth.
Test one prompt pair
Observe how minimal language shifts activate different semantic attractors and response trajectories.
Study one naming system
Ask when a term functions as ornament, label, or real semantic infrastructure.