Context Precedence
[Established]
Operational. The tokens a model sees first exert disproportionate influence over everything that follows. A session is not a neutral container filled equally by each turn; the opening context sets the dominant pattern the rest of the session extends. Loading high-density, coherent context first therefore shapes the entire trajectory more than any later correction can.
What is known: early-context dominance is a documented, reproducible property of trained language models. What is open: why it holds as strongly as it does, and what exactly makes one opening more dominant than another, is not fully characterized — that is the question the rest of this taxonomy circles.
Metaphor (grounded + bounded). “The first tokens have the strongest gravitational pull.” Grounded: later tokens are conditioned on earlier ones, so early structure propagates. Bounded: there is no literal force; the pull is statistical conditioning, and a later turn can still redirect a session — it just works against the gradient rather than with it.
Coherence Triggers Compression — “Notes cohere in symphonies; everything else is noise”
[Hypothesis]
Operational. The working axiom of the discipline. A model carries, from training, dense latent structure — “compressions” of concepts and their relationships. A coherent opening context surfaces that latent compression and lets the session operate from it. An incoherent or absent opening produces only the model’s best-effort fluency: articulate, internally consistent, but ungrounded. The claim is not that coherence adds knowledge; it is that coherence activates structure already present.
What is known: dense, coherent preseeds produce qualitatively more grounded, drift-resistant sessions than sparse or incoherent ones — reproducibly. What is open: whether “coherence triggers compression” is the correct mechanistic account, or a useful description of a behavior whose cause is something else.
Metaphor (grounded + bounded). The orchestra and the conductor. Grounded: the latent capability (“orchestra”) is present regardless; the preseed (“downbeat”) determines which score gets played. Bounded: there is no agent conducting and no music — the figure describes selection among already-present capabilities, not the summoning of new ones.