DR-010 — MemPalace is a recall feed, not the decision register

The canonical store for binding decisions is THIS markdown register — durable, curated, git-diffable, citable (DR-NNN), and it survives MemPalace outages (the palace was unreadable for ~2 weeks in Jun 2026). MemPalace is a derived, lossy recall layer (auto-mined, noisy, no supersession) that feeds the register via mining; it does not replace it. Minimize duplication by linking a DR to its source drawer/session, not copying the evidence. Cited as binding by runbooks/mempalace-repair.md.

(Swept unrecorded at ec9fcfde; re-accepted by owner ruling S499 — the S499 tombstone audit found it still binding with no other home.)

Amendment (S517): the register/palace split holds, and S517 measured both halves. The negative claim is now evidenced: asked whether the OKF derivation-vs-canonical-store fork was settled, vector search ranked a 1 July proposal (“answers the fork without a vision rewrite”) above the ratified outcome — DR-023/DR-025, two days later, went the other way. Nothing in the result marked it superseded; only authored_at plus the register revealed it. The palace has no supersession and cannot acquire one by mining.

The positive claim is the amendment: the palace is the only home for what was argued, doubted and refuted. S517’s two most load-bearing findings — that fork, and the still-live contradiction in initiatives/core-product/canonical-pipeline/intended-architecture/02-data-flow.md S2.1 (which quotes a sentence 01-vision.md records as retired by DR-025) — came from subagent transcripts, not from any doc or DR. So the routing is three-way, owner-ratified S517: register = what was decided; ordna task files + git = when; MemPalace = what was argued. Do not prune the palace toward “only decisions” — that deletes the one thing it uniquely holds.