Left: a default pipeline moves content. Nothing in the message says which surface it came from, so agent C can quote agent A's summary as an outward fact - the section 5 shape of the Globe and Mail claim. Right: the same hand-offs carry shell type, pair id, and proof tier, so refusal and documented absence are available moves.
Adding agents lengthens the strip; it does not add layers. Shared context is a longer window with the same decay, which is why the third strip is wider than the first but built from the same material. The semiosphere strip is stacked instead: fond, shell position, pair-read warrant, an inward loop that never crosses the filter, and dated ingress that retools operators without raising volume.
Left: both agents run one model, one tokenizer, one geometry. Their exchange moves within the manifold, and topic is the only organiser - the condition section 6 measured as theme gravity. Right: a bilingual filter is partial. Some material crosses recoded, some is refused, and some expected outward sentence is simply absent, which is a finding rather than a gap to fill.
| Figure | Section 10 claim | Montreal anchor |
|---|---|---|
| M1 | Agents are animators, not shells; typing has to be engineered into the channel | §6 flat vs shell-scoped retrieval; §5 unpaired claim |
| M2 | Shared memory is a longer window, not non-hereditary memory | §4 operator change without volume change (1983–87) |
| M3 | Agent paraphrase is primary-system coupling, not bilingual relay | §1 retracted rotation; §2 language confound; measured round trips MA12, journal §14 |