Glossary

One append-only log is the source of truth and every query shape — KV, document, BM25, vector, hybrid, memory — is a view over it, so views can never drift and time-travel is free.

Does not promise Not a distributed log. A single node owns the log; replication is Raft, and Raft is simulation-tested, not production-proven.

A density ladder of selectable vector compression tiers over one index: Prism-8 (int8, ~8x, default), Prism-1 (binary, ~32x, deprecated), Prism-R (RaBitQ, error-bounded, graph-integrated).

Does not promise The tier is a space/recall trade, not free. Prism-1 measured recall@10 ~0.10 and is deprecated; only Prism-8 and Prism-R are recommended.

The sub-bit tier of the Prism ladder: k-means residual coding against the Beacon centroid, orthonormal projection to dim/shrink, then a 1-bit sign code with unbiased estimation, corrected by Facet rerank. Measured at 768-dim: recall@10 0.9950 at 57.5 B/vector.

Does not promise That figure is from a synthetic clustered corpus. Real-embedding validation is pending. The shrink=4 tie is flagged as likely rerank-tier absorption and is not trusted yet.

An in-memory centroid graph (repurposed HNSW) routes queries to disk-resident quantized postings — SPANN/SPFresh-style ANN that reads a bounded fraction of the corpus rather than all of it.

Does not promise Bounded reads are a routing property, not a latency SLA. Recall depends on nprobe, and the disk tier inherits the storage engine’s own latency.

Filtered ANN of ACORN lineage: the predicate is evaluated during neighbour expansion instead of pre-cutting the graph, so a tight filter does not collapse recall. A filtered-out node still serves as a stepping stone.

Does not promise Does not make every selectivity equal. Very selective filters deliberately fall back to Sieve, an exact scan of the eligible set.

A Prism-R index’s rotation is reconstructed from a stored seed rather than the matrix: because the orthonormal rotation is deterministic in (dim, seed), a snapshot persists { seed, dim, 1-bit codes } and boots from a compact fingerprint.

Does not promise Not a portable format guarantee. A seed mismatch is refused outright, because the wrong seed yields garbage codes.

The persisted checkpoint chain can never resurrect restored-away data nor silently drop un-checkpointed writes, across a rewind or a plain restart. Checkpoint artifacts are deleted before the WAL is truncated, a monotonic checkpoint epoch fences detached writer threads, and open re-seeds dirty tracking from the replayed WAL tail.

Does not promise A durability guarantee about the checkpoint chain, not about the disk. It does not protect against media failure, and it is not a backup.

Structural tenant isolation on two fences. Every tenant entity is kind-prefixed t_{pid}___, so a credential addresses only its own compartment; and every credential carries a platform_identity bool checked before any role or uid comparison, so a project-plane credential can never assert platform authority.

Does not promise Isolation is namespace-based, not cryptographic. Tenants share one process, one log, and one page cache — this is not a per-tenant encryption boundary.

Tenant data portability with a sensitivity boundary distinct from the isolation boundary: the namespace exports as integrity-hashed NDJSON from one pinned snapshot, and the tenant’s own auth material is redacted on the way out.

Does not promise Not a full backup. auth_otp and auth_refresh are excluded outright and password_hash is emptied, so an export cannot restore a tenant’s sessions.

The disk tier stops being paid for in RAM: settled SPANN postings are served zero-copy from mmap at query time, so the page cache is the tier and the heap holds only the navigator, centroids, per-posting headers and the live memtable.

Does not promise Resident memory drops; total I/O does not. The bytes still get read — they are just the kernel’s problem, so behaviour depends on page-cache pressure.

Tenant decommissioning that states exactly what it deletes and no more. A purge tombstones every t_{pid}___* entity, destroys the tenant’s keys, then compacts — idempotent and resumable, with an honest live recount deciding completion rather than the loop’s own counter.

Does not promise This is tombstone-plus-compaction, NOT crypto-erasure. The declared erasure_scope is "live checkpoint + retained WAL"; bytes survive in the retained WAL, restore-undo backups, replica logs, and Disk-profile segments pending LSM compaction.

A tenant-scoped read costs O(that tenant), not O(every tenant). A per-kind id posting index answers a bare Filter::Kind directly instead of falling through to a full-corpus scan.

Does not promise Default-on for the Ram profile only. The Disk profile ships default-off pending burn-in, and an unflagged root merely scans — correct, only slow.

Test the path production actually takes, and make sure that test can fail. SPANN filtered search dispatches on filter size, and the large-filter branch — the only one production reaches — was the one without coverage.

Does not promise A testing discipline, not a runtime feature. It says the measured path is the real one; it does not say every path is covered.

An index that may over-approximate is only safe if every consumer re-checks the full predicate, so the re-check obligation travels with the candidate set instead of living in a convention. A filter resolves to candidates tagged exact or approximate, and the obligation is discharged when owed.

Does not promise Not a tenant-leak guarantee under attribute-based tenancy. It is load-bearing precisely for callers that encode tenancy in an attribute rather than in the kind.

An index read refuses to answer while the projection that feeds it is behind the log. Every persisted index arm checks one shared, single-sourced predicate before trusting its postings, and otherwise falls back to the exact scan.

Does not promise Availability is deliberately traded for correctness: in the lagging window reads get slower (exact scan), not wrong. Over-approximation is survivable; under-approximation is not, so the fallback is mandatory.