DSET-Q35: Strong-MoE Script Capacity Ceiling

Status: frozen read-only development ceiling, 2026-08-09. No Qwen3.6 output exists at freeze time. No training or holdout is authorized by this contract.

Question

DSET1 on OLMoE is already causal and strong on numeric edits (98.05%) but weak on choice edits (71.48%). PSET1 shows that replacing the host LM with a small isolated value decoder destroys corrected-value synthesis. DSET-Q35 asks whether a current stronger MoE can follow the exact same model-owned edit protocol without training.

Frozen implementation

• host: local pinned Qwen3.6-35B-A3B MoE artifact;
• data: exactly the 256 already-opened PSET1 diagnostic identities, mapped back to their immutable DSET1 clean/fault presentations;
• prompt, script grammar, deterministic executor, maximum 32 generated tokens, greedy decoding, and exact complete-trajectory scoring unchanged from DSET1;
• no demonstrations, prompt variants, verifier, repair, router, tool, answer label, adapter, or benchmark-specific selection;
• eight disjoint H100 shards, merged before inspection.

The ceiling passes only if exact execution is >=95% overall, exact scripts are >=90% in both numeric and choice families, clean copy is >=99%, fault repair is >=90%, and there are zero execution errors or exhausted rows. A miss closes this exact untrained transfer prompt. A pass authorizes a separate prospective trained DSET transfer contract; it is not itself a reasoning claim.