Skip to main content
Imagine an answer begins, “The invoice should be classified as…” and the provider connection then breaks. A second model might produce a plausible continuation, but it would not be the same generation. Gateway’s design refuses to present those two outputs as one uninterrupted answer. The important boundary is when meaningful output may have been exposed, not whether Gateway can prove that a person read it.

Record before releasing

For incremental delivery, Gateway holds the first meaningful fragment while it records and confirms the output marker. Text, tool-call arguments and refusal content cross this boundary. Liveness, framing and reasoning-size observations do not establish visible answer content by themselves. The marker is deliberately conservative. If the connection fails immediately after release, Gateway may not know whether the caller received the fragment. It must still preserve the possibility of exposure.

Before and after are different recovery conditions

Before commitment, the live execution owner may use a policy-permitted recovery action under the original plan and remaining limits. That permission still depends on failure kind and certainty; it is not automatic. After commitment, Gateway may continue delivery from that attempt or report a truthful partial or uncertain ending. It cannot switch provider, splice in a new model’s answer or reset the call budget. A caller can explicitly authorize a new inference. That produces separately identified work, not an invisible continuation of the old stream.

Buffered results have a different release gate

A buffered answer is not exposed incrementally. Successful content is held until the required final persistence is confirmed. A validated document is also held until its output contract passes. This does not mean every no-output failure needs an invented stream marker. A failed, cancelled or uncertain call can legitimately finalize without output when its complete accepted record shows that no output fact is required. The rule is to enforce the actual mode’s release boundary, not add a marker merely to make every path look alike.

Exposure, provider completion and delivery remain distinct

The provider may complete while the caller later stalls. Gateway must retain the known provider result and the separate delivery failure. If the caller stalls while the provider outcome is still unestablished, the delivery failure does not prove the provider completed. A local output marker also does not prove that the caller saved the answer for replay. The benefit is answer coherence and honest attribution. The trade-off is an extra protected boundary before visible output—and explicit partial failure when continuation would be misleading. See Following a live stream for the full sequence, and Execution evidence for what the marker records.