LM Studio divides the context window among parallel slots

This field note comes out of running Tycho's grading pipeline — the stage that builds The Ghost — on local open-weight models.

Loading a 27B model with -c 16384 --parallel 4 doesn't give 16k context per request — it divides the budget among slots, so each gets about 4k. That's under half what a grading prompt with few-shot examples needs. A four-item probe failed all four with Context size has been exceeded before any inference ran.

The fix: -c 32768 --parallel 2 on 32 GB gave 16k per slot. Halving parallelism to double per-request context solved it.

Document the arithmetic. Budget context per slot and write it down where the next person loads the model — this one went into the repo's instructions the same day. A harness that "makes real inference calls" and one that "fails before inference" produce the same red test; read the error string before concluding anything about the model.

Where this goes next

Run a small parallel probe first to catch misconfiguration before a long pass burns time. Working means a probe that fails on the model, not before it.