CSM edges Hindsight on BEAM 100K

Promising, but still a single-trial, local artifact comparison. The result is interesting because it shows a shard-based memory pipeline can trade latency for better BEAM 100K accuracy and tighter answer context, but the scientific bar is not “higher score once.”

• –Run repeated trials and report variance or confidence intervals. The repo itself notes single-trial behavior and temperature-0 nondeterminism, so the current delta could move on rerun.
• –Publish a preregistered protocol and keep tuning off the test set. If shard counts, retrieval settings, or prompt details were adjusted after peeking at BEAM, the comparison risks benchmark overfitting.
• –Make the cost story symmetric. CSM reports internal shard-token spend; the accepted Hindsight artifact does not, so a fair efficiency comparison needs matched accounting or an explicit limitation.
• –Add independent replication on different hardware and, ideally, official chart acceptance. That is the difference between a credible local result and a durable benchmark claim.