A MySQL connection can succeed while the answer depends on whichever session the pool handed to the AI.
One connection uses UTC and strict SQL mode.
Another inherited a local time zone, permissive grouping rules, or a different collation.
The SQL executes in both sessions. The meaning is no longer the same.
Treat the session as a versioned runtime contract:
- approved database and authenticated role
- time zone and reporting cutoff
- SQL mode
- character set and collation
- transaction isolation and read-only state
- row, time, and result-size limits
Initialize those values outside model control, verify the effective session before use, and attach a normalized contract fingerprint to the operation trace.
For pooled connections, rollback open transactions, clear workflow-owned state, reapply the baseline, and reject the checkout when verification fails.
Then test pairs that should agree: new versus reused connections, changed server defaults, strict versus missing SQL modes, international text comparisons, and a prior request that deliberately dirties the session.
“The query ran” is not the acceptance criterion. The same governed question should preserve the same execution semantics.
Full guide: How to connect MySQL to ChatGPT with a session contract
Top comments (0)