SOTA security experiment: private outsourced vector retrieval with compact codes

Finding

Spruce, arXiv:2609.03376, co-designs compact binary retrieval codes with two server MPC for private outsourced dense retrieval. Across 383K to 5.42M document corpora, the originating team reports full scans of 0.21 to 2.97 seconds at 10 Gbps, private pruning of 0.06 to 1.09 seconds, 13.1x to 22.9x speedup over the closest measured prior work, and 93.9 to 97.3 percent of full float NDCG under pruning.

Evidence class: originating team measured. Not independently reproduced by RuVector.

Opportunity Score: 4.464.

RuV mapping

RuVector, Cognitum sovereign retrieval, Core Memory enterprise memory, RVM privacy policy, RVF provenance, MCP retrieval tools, and distributed infrastructure.

The strategic primitive is private candidate generation over an untrusted retrieval service, not a wholesale replacement of RuVector search.

Reproduction first experiment

Compare:

1. ordinary RuVector retrieval inside a trusted boundary
2. encrypted transport to an untrusted retrieval service without query privacy
3. compact binary candidate generation under a two server non collusion model
4. private cluster pruning only if the full scan baseline is reproduced

Pin corpus, embeddings, binary code model, reranker, network bandwidth and RTT, CPU model, memory, server topology, seeds, query set, and cryptographic library versions.

Metrics

NDCG, recall at 10, candidate set size, end to end latency, throughput, communication bytes per query, CPU time, memory, preprocessing time, owner dealer load if used, infrastructure cost, and privacy threat coverage.

Security falsification

The non collusion assumption is load bearing. Model compromise of both servers, traffic correlation, malicious server behavior, codebook leakage, query repetition, tenant mixing, and denial of service must be analyzed separately. Do not market the mechanism as private if the deployment topology cannot enforce or independently operate the two trust domains.

Promotion gate

At least 90 percent of full float NDCG, p95 latency under 3 seconds on a million scale corpus under the pinned network envelope, communication at least 10x below a naive secure baseline, zero plaintext query or corpus embeddings visible to either single untrusted server, and a documented rollback to trusted boundary retrieval.

Governance

Research branch only until the threat model and legal or customer privacy requirement justify the complexity. No custom cryptographic primitive. Use audited standard MPC and OT libraries only. No autonomous deployment or tenant migration.

No implementation file or test entry point is named. Start by defining the four retrieval comparisons and pinning the corpus, embeddings, network, hardware, topology, seeds, query set, and library versions. Done means the listed retrieval and security metrics meet the promotion gate, with a documented rollback and separate analysis of the stated threat cases.