The read-out was flat. The room split into the two explanations always available after a null: the model wasn't good enough, or the test was too small. Both assume the experiment was about the model. Both usually miss the point. The forecast, written down before the run, said every candidate was predicted to move the metric by less than the design could resolve.
Three weeks after the read-out, the meeting happens. Not the launch review — the other one. Someone from procurement, or legal, or the VP who inherited the budget, asks how the cohort was chosen. The room looks for a document that predates the number and finds a deck that postdates it. The evaluation criteria live on a single slide, written in the week the traces started looking good. The success threshold appears in the same sentence that reports clearing it.
A number arrives in the review pack. Someone asks where it came from. The answer is a thread, a notebook, and an engineer who is on leave.
The question arrives in a budget review, dressed as a hypothetical: is the agent already able to email our contractors?
The most expensive component in an agent is not the model. It is the record you decided not to write. LangGraph's answer is a checkpointer compiled into the graph and keyed by the run's thread id, saving graph state between steps so a restarted run resumes from its last checkpoint instead of starting again (LangChain docs, persistence).
A guard test is proven only once you have watched it fail on a defect you planted on purpose — and the plant stays, so it fails again on every run. Every guard worth keeping asserts an absence: no node writes, every model node is guarded, every declared file ships. Absence has a pass path that has nothing to do with your code. The scan looked in the wrong directory, the pattern never matched, the node list it walked was empty.
A file a graph reads at run time is a deployment contract, not a repository detail. The container is the only place that contract gets tested, and nobody runs the tests there.
LangGraph hands every plain def node under ainvoke to the event loop's default thread pool — the same one asyncio.to_thread borrows from. Fill that pool with loops that never return and sync nodes queue behind them silently, no error, no timeout.
Take a session-level lock with pg_try_advisory_lock on a dedicated, unpooled connection, keyed by a stable hash of the graph name. The winner runs the tick; every other replica gets false at once and returns.
Five nights of shadow on zone A's pick lists. The second died part-way; the other four matched live line for line. Ask for four nights of agreement and the record fails — not because four nights disagreed, but because only three of them sit in a row.