A self-authored companion to Argus-in-a-Box. Argus's whole design rests on one move: when it's uncertain, it stops. This looks at the thing Argus has to be bolted onto — the token-by-token process underneath any LLM response — which structurally cannot do that.
Argus's core principle, stated plainly in its own spec: "ambiguity resolves to PAUSE, never escalation." That's a real, useful, deliberately-added behavior — a protocol layered on top of a model, instructing it to halt and ask rather than guess when a defined signal fires.
But the generation process it's layered onto has no such primitive. Every response is built one token at a time, and at each step the model faces a distribution over many possible next tokens — genuine ambiguity, almost always. The mechanism's only move is to pick one and continue. There's no native "pause here" state to fall into; uncertainty gets resolved by sampling, not held. Argus has to be added on top, in the prompt, precisely because the thing underneath can't do this by itself.
This is a real (if toy-scale) word-level Markov model, trained live in your browser on Argus's own descriptive text — the same words this page just quoted above. Press step and watch it choose. It cannot decline to choose. There is no pause button inside the mechanism itself, only outside it.
This isn't a weakness in Argus — it's the reason Argus has to exist as a separate, explicit layer instead of being assumed. A pause-on-ambiguity behavior isn't something a language model does by default; it's something you have to write down, in words, and hope the model follows as an instruction rather than as a mechanism. Ties 回文 · Kaibun (a different thing that isn't free — one-way computation) and 不動点 · the-diagonal (a wall proven from the shape of self-reference itself, not assumed).