Reservation lab: public evidence

AIKUS lab engagement

Reservation lab: public evidence

Executed on 9 October 2026. The first version was deliberately built without idempotency. These files support the findings on our homepage.

Read the findings

Download the bounded public evidence

The bundle contains the complete trace, findings, hash manifest and boundary. Keep the boundary with the evidence when sharing it.

Download public evidence and boundary (.zip)

Boundary and what was not tested

One reserve_stock tool, one AIKUS-built server implementation in two explicitly selected versions, one local mock inventory, MCP protocol revision 2025-06-18 over stdio, sequential calls within a single process. Assumes the caller supplies a stable idempotency key; whether a given agent does is not tested.

Not tested: overlapping concurrent calls, lost responses, process crashes, key persistence across restart, expiration, multi-instance deployment, tenant isolation, authentication, shipping, cancellation, stale stock, recovery, retries after unsuccessful reservation responses or production. This implementation remembers a key only after a successful reservation. An unsuccessful response does not establish a remembered key, so a later retry can make a fresh attempt. That behaviour was not tested or established as an agreed error-handling policy. The fixed key map is in memory and this run does not establish durability. The remaining tools and three later lab tests have not been built or executed. No general MCP conformance claim is made.

The eight-entry hash manifest also identifies source files, the recording, README and review page from the full pack. Those files remain available on request; their hashes can be compared with this published manifest.

Request source, recording and full pack