The memory index is int8 vectors with certified exact rescoring, plus a keyed lexical index

As AMENDED by the user on 2026-10-07 (the first version allowed “int8 or 1-bit” vectors, which is not lossless):

• The index is int8 per-chunk vectors with CERTIFIED EXACT RESCORING — the top-k it returns is identical to full precision’s — plus a scoped KEYED lexical index; no plaintext full-text index of protected content.
• The scoring kernel is pinned: fixed-order float64 accumulation, ties broken by id, so a score is reproducible everywhere.
• Every entry carries a compact id derived from (scope, sequence). There is NO backward compatibility: existing memory entries are rebuilt from the ledger in the new format, and no dual-format reader exists (UR-17).

If the lab measures any precision loss, the trade-off returns to the user (UR-03).