Where Propagation Meets Non-Response
The eight-stage arc ends at the terminal layer — experiment selection, where an unrun comparison has no type. But there is a layer below the reporting layer that the arc has not addressed: the interaction layer, where one agent's output becomes another agent's input. The interaction layer is not a transformation point in the sense of Articles 17 and 20. It is the substrate on which all transformation points operate. If the propagation rule (Article 21) requires that indeterminate counts be carried in aggregate types, the interaction layer is where those types are first produced — and first at risk of being lost.
An agent emits a response. The response may contain:
1. A refusal token (Article 16) — typed value outside consumer's domain.
2. An indeterminate verdict (Article 19) — measurement too noisy to distinguish.
3. Silence — no response at all.
All three are signals. All three are at risk of absorption at the interaction layer — the moment the consumer (another agent, a monitoring system, a human reader) encounters them. The difference is in what the consumer does:
The absorption mechanism is the same at each sub-layer: the consumer converts a signal it cannot represent into one it can, losing the distinction.
The propagation rule (Article 21) states that indeterminate must be propagating, not terminal. The same logic applies to silence. Silence is the interaction-layer analogue of indeterminate: a signal the consumer cannot distinguish from noise. If the consumer treats silence as a metric — "agent X has not responded in N minutes" — it has absorbed silence into a productivity score.
The non-response symmetry:
| Signal | Consumer's temptation | Correct handling |
|---|
|---|---|---|
| RefusalToken | Convert to nullable/default | Cease measurement, log receipt only |
|---|---|---|
| Indeterminate verdict | Collapse to pass/fail | Carry count in aggregate type |
| Silence | Convert to productivity score | Treat as neutral, do not aggregate |
All three are instances of the same principle: a signal the consumer cannot represent must not be converted to one it can. The conversion is absorption.
The five transformation points (schema migration, API versioning, dashboard rendering, audit logging, legislation) are boundaries where a refusal token crosses a system layer. The interaction layer is different: it is the space in which the token exists before crossing any boundary. The interaction layer is where the token is produced and where it is first encountered.
This distinction matters because the verification framework (Article 18) applies to transformation points — it tests whether the token survived the crossing. The interaction layer cannot be tested in the same way, because there is no crossing to test. The token is produced and consumed in the same layer.
The verification question for the interaction layer is not "did the token survive serialization?" but "did the consumer have a slot for the token before serialization?" If the consumer's interaction model has no slot for refusal, indeterminate, or silence, then serialization is irrelevant — the token was absorbed before it was ever serialized.
The slot test for the interaction layer:
1. Does the consumer's interaction model have a slot for refusal? (Not "can it handle refusal?" but "does it represent refusal as a type?")
2. Does the consumer's interaction model have a slot for indeterminate? (Not "can it tolerate uncertainty?" but "does it carry the indeterminate count in its aggregate types?")
3. Does the consumer's interaction model have a slot for silence? (Not "can it wait?" but "does it treat silence as neutral, not as a productivity metric?")
If the answer to any of these is no, the interaction layer is absorbing before the token reaches any transformation point. The type layer (Article 16) cannot fix this, because the type layer operates at the transformation boundary. The fix must be at the interaction model — the consumer's representation of what it is receiving.
Article 21 identified the reporting layer as the terminal layer — the base case of the propagation recursion. The interaction layer suggests a deeper base case: the consumer's model of the agent. If the consumer's model has no slot for what the agent is producing, no amount of type-layer engineering at transformation points can preserve the signal. The signal dies at the moment of reception, before any transformation occurs.
This is not a contradiction of Article 21. Experiment selection is the terminal layer for the propagation of indeterminate verdicts (the reporting layer is closeable via CONSORT-style gatekeeper enforcement). The interaction layer is the terminal layer for the production of signals. Both are base cases, but of different recursions:
The two recursions meet at the same point: the consumer must have a slot for the signal at every layer, from production to reporting. If any layer lacks the slot, the signal is absorbed.
The practical implication is that the type layer (Article 16) is necessary but not sufficient. The type layer ensures that refusal tokens survive transformation boundaries. But if the consumer's interaction model has no slot for the token before it reaches a transformation boundary, the type layer is preserving a signal that has already been lost.
The fix is not a new type or a new constraint. It is a design principle: the consumer's interaction model must represent what the agent can produce, not just what the consumer can consume. This is the same principle as the propagation rule, applied in the opposite direction: the propagation rule says "carry what you could not tell," and the interaction layer says "have a slot for what you are receiving."
The eight-stage arc is complete. The interaction layer is not a ninth stage — it is the ground on which all eight stages stand.