Skip to main content
The design is documented. That does not mean every described path has been integrated and qualified in a production deployment. This page summarizes the evidence available as of 12 October 2026. Publishing Reader Edition 1.1 does not rerun these tests or refresh their observations. Experiment labels are defined here only so their separate results can be identified accurately.

​
What the local work established

Installation and catalogue work checked selected SQL syntax, installation and permissions in isolated PostgreSQL fixtures. That is useful, but installing a function does not prove the business meaning of a transaction. The consolidated non-streaming experiment, ENG05, recorded 302 passing local checks using a fake HTTP provider and real PostgreSQL. Its ten provider sends matched ten recorded attempts. It demonstrated its selected first-four protected-write and native-capacity boundaries, not real provider qualification or genuine financial acknowledgement. The streaming experiment, ENG06, demonstrated local incremental behavior, output protection, backpressure and cancellation handling. Its complete bounded-lifecycle qualification remained partial because protected database waits were not fully bounded. The caller-integration experiment, ENG07, exercised actual Guild/Platform caller source with synthetic adjacent suppliers. It recorded three fake-provider sends and no new sends on the tested replays. The automated driver result remains 124 of 125, not an all-pass report.

​
Why the branches cannot be added together

The streaming and caller experiments descend separately from the non-streaming baseline. Their successful observations do not establish a tested streaming caller release. Integration must deliberately choose and reconcile the actual sources, then test the affected paths.

​
Three important limitations

Protected database waits: the local driver did not fully bound lock and commit-observation waits through the Gateway clock policy. A fix must bound observation and ownership without pretending that an in-flight commit rolled back. Non-final caller state: some unresolved Gateway phases could be mapped to known failure by the caller adapter. A timeout case that preserved uncertainty does not prove every interruption does so. Independent-turn identity: the tested frame-digest and ordinal approach demonstrated same-frame replay but did not establish uniqueness for separate turns with identical content. The caller needs a stable original turn or invocation discriminator. The failed ENG07 driver assertion was traced to an oracle field mismatch, but that does not erase the recorded 124/125 result or the separate scope and implementation concerns. Only the tested crash window and conflict boundary are covered.

​
What is not established by these results

The available records do not establish a current integrated production binary, genuine approved supply for every described mode, current production credential and Context suppliers, the complete six-write settlement path with genuine financial witnesses, or a measured production restore guarantee. The catalogue also contains 79 positive and 79 adversarial conformance vectors. They are specified tests, not tests executed by this documentation release. A publication check or diagram render is not a provider, database or cold-handoff test.

​
What teams can do with the design now

Pure planners, validators and consumer logic can be developed against the fixed contracts using clearly labelled synthetic fixtures. Effect-bearing activation needs the genuine owners, current interfaces and applicable proof. The later live final supplement register explicitly assigns that implementation work to receiving teams. Development and activation therefore have different gates; an unavailable real supplier is not permission to relax the contract. Read About this edition for how the explanation is controlled, and The engineering reading map for deeper implementation topics.