Production AI evaluation: which failures should block a release?
A better average score can conceal a broken requirement. Define release criteria for a document-grounded assistant, its graders, and the evidence behind approval.
Read the noteThe work behind the standard.
Notes on the decisions, mechanisms, and tests behind consequential software.
Deployment health: prove a business operation before moving traffic
A passing health endpoint can miss a broken workflow. Define a bounded release probe that follows one accepted operation to its result and leaves enough evidence to investigate uncertainty.
Read the noteArchitecture assessment: what would justify a rewrite?
Turn a repair-or-rewrite debate into a bounded investigation. Define the evidence that would change the decision before commissioning another system.
Read the notePgBouncer rolling restarts: why idle clients delay exit
A local PgBouncer 1.26.0 study shows why a committed transaction can leave the process waiting for an idle client to disconnect.
Read the notePostgreSQL migration rollback: find the first incompatible write
An additive schema change can still make rollback unsafe. Use a scheduled-job example to plan compatible releases and test recovery after new data arrives.
Read the noteDurable transaction recovery after a lost response
A Rust reference study of durable receipt recovery when a transfer commits but its response is lost. Inspect the failure experiments and operating limits.
Read the case studyAPI idempotency: what survives a retry?
How command identity and recorded outcomes make API retries safe. Inspect a Rust reference study and the persistence boundaries it does not establish.
Read the note