A portable, model-agnostic protocol: it notices when an AI stops making real progress and forces a pause so a human can re-aim it. That's the whole job. It does not decide what's correct — it decides when nothing may be done until a human re-authorizes action.
This runs the protocol's own threshold policy client-side: W0/W2/W3/W5 are immediate-pause signals; W1/W4/W6/W7 need two checked before pausing is warranted. Nothing here is judging your actual conversation — it's just the documented rule, made checkable.
Paste the block above at the start of a session, or right before recursive refinement begins. Works in Gemini, ChatGPT, Claude, or any LLM — it's a protocol, not a platform feature. Only caveat: some models (Gemini especially) are more willing to keep going unless the pause phrase is explicit, so use the literal trigger ARGUS PAUSE — HUMAN DIRECTION REQUIRED and require the model to stop output after the pause report.
Resume only after you provide: a confirmed/revised goal statement, at least one measurable success criterion, and an approved next-step approach.
=== END PAUSE REPORT === must be the last output, no trailing text) so a model can't keep "helpfully" continuing after it's supposed to have stopped.Three documented iterations — the weight carried into az1's asteroid belt.
Argus's core move is ambiguity resolves to pause, never escalation. But the substrate underneath any LLM response — token-by-token generation — has no such primitive: every token is a forced choice, resolved by sampling, never by waiting. 止まれない · Tomarenai ("cannot stop") builds a small live token-sampler to show the thing Argus has to be bolted on top of, because it can't do this by itself.