RFD 1066: Differential Mamba for caption encoding

Problem

RFD 1065 resolves a dataset caption to a trait capability id through HRR/HRR.Cleanup. A question came up: does Differential Mamba (Schneider, Zimerman, and Nachmani, arXiv:2507.06204), a sequence model architecture, improve that step.

Decision

Do not adopt it. Abandon this line of work.

Differential Mamba reduces attention overallocation in a trained, autoregressive language model, for long-context retrieval. RFD 1065’s resolve step needs none of that. HRR.encode_atom/2 is a closed-form hash into a phase vector, with no training step and no sequence to attend over. HRR.Cleanup.nearest_above/3 compares that vector against a codebook by cosine similarity.

RFD 1021 already chose this closed-form algebra over a trained model, for the same reason a fact vector needs no LLM. A sequence model here would trade an exact, cheap, verified mechanism for a trained one. That trained model solves a problem this step does not have.

Related

RFD 1065 owns the resolve step this RFD examined. RFD 1021 gives the HRR library and the earlier closed-form-over-trained-model choice.