Invitation: two week review window before MemPalace's row goes on the leaderboard

I maintain agent-memory-bench, an execution-graded benchmark for memory layers in coding agents. MemPalace is an arm. I also wrote RE-call, which is another arm.

The offer, first. An eight-arm run lands within hours, and the leaderboard has published nothing yet. I will not put MemPalace's row on it until you have had the full two week review window, or until you tell me to go ahead sooner. The board is the canonical record here, so the window happens before the record that counts.

What I want you to reject. Every judgement call is stated in adapters/mempalace/VENDOR_REVIEW.md. The four that move your numbers: ingest runs --mode convos rather than our markdown render; twenty of your forty-four tools are exposed and every write tool is withheld (the corpus is frozen after ingest, so a write would change the store the next seed reads; the same applies to RE-call); --max-chunks-per-file sits at your default because setting it would be me tuning your product; and the palace path is refused above 120 characters, working around an onnxruntime DLL failure that chromadb re-raises as "the onnxruntime python package is not installed" when it is installed.

One finding that runs against me. Your store holds ~1,590 items from the same ~4,900 documents RE-call chunks into 28,000 to 55,000. That is your consolidation working as designed. It also means a planted document lost at ingest is scored by me as a retrieval miss when it is a write-path miss, and nothing in the design separates those. That would deflate MemPalace specifically. It is untested and I have logged it as an open problem, not a result.

Two offers: sponsored credits for your arm's traffic, with execution staying on the shared harness and the artifact recording who funded what; and a vendor-tuned variant run as an additional labelled arm. Not on offer: vendor-executed cells in the canonical table.

Every per-session record is published and the configs are sha256 pinned. If my arithmetic is wrong, python -m scripts.verify_run --all proves it with no credentials, no database and no network.

Wired against mempalace==3.8.0.