Skip to content

research(nightly): semantic-drift-detector — PSI + coherence drift detection for agent memory#551

Draft
ruvnet wants to merge 1 commit into
mainfrom
research/nightly/2026-06-11-semantic-drift-detector
Draft

research(nightly): semantic-drift-detector — PSI + coherence drift detection for agent memory#551
ruvnet wants to merge 1 commit into
mainfrom
research/nightly/2026-06-11-semantic-drift-detector

Conversation

@ruvnet

@ruvnet ruvnet commented Jun 11, 2026

Copy link
Copy Markdown
Owner

Summary

Adds nightly RuVector research for semantic-drift-detector — a pure-Rust semantic drift detection library for agent memory vector collections.

  • Implements three composable DriftDetector variants under a shared trait: CentroidDrift (L2 centroid shift), PsiDrift (Population Stability Index on cosine histograms), CoherenceDrift (PSI + intra-window coherence delta)
  • Zero external library dependencies; rand only for test/benchmark data generation
  • All 13 tests pass; benchmark ACCEPTANCE RESULT: PASS (9/9 variant–dataset combinations, TPR=1.00, FPR=0.00)

Benchmark evidence (cargo run --release -p ruvector-drift --bin benchmark)

Hardware: x86-64, Intel Celeron N4020, Linux 6.18, rustc 1.94.1

Variant N D mean_us p50_us p95_us TPR FPR Acceptance
CentroidDrift 5000 128 1046 982 1338 1.00 0.00 PASS
PsiDrift 5000 128 2601 2556 2931 1.00 0.00 PASS
CoherenceDrift 5000 128 5813 5782 6946 1.00 0.00 PASS

Deliverables

  • Crate: crates/ruvector-drift/
  • ADR: docs/adr/ADR-199-semantic-drift-detector.md
  • Research doc: docs/research/nightly/2026-06-11-semantic-drift-detector/README.md
  • SEO gist: docs/research/nightly/2026-06-11-semantic-drift-detector/gist.md

Ecosystem connections

This PoC connects: RuVector vector search · ruvector-mincut (graph-cut compaction) · ruvector-temporal-tensor (window storage) · ruvector-verified (proof-gated re-embedding) · ruFlo (observe→compact→reindex workflow) · mcp-gate (agent introspection tools)

What's not claimed

  • TPR=1.00/FPR=0.00 is on large synthetic drift (mean shift 2.0). Real workloads need calibration.
  • No comparison to Milvus/Qdrant/etc. — no public drift detection benchmark exists for vector databases.
  • CoherenceDrift's O(N²) coherence is sub-sampled to 200 vectors; introduces ~5–10% estimation error at N=5K.

Next steps (not in this PR)

  • Incremental streaming detector (O(D) per vector)
  • ruvector-drift-wasm WASM build
  • MCP tool surface in mcp-gate
  • ruFlo trigger schema

Generated by Claude Code

Adds ruvector-drift: a pure-Rust semantic drift detection library for
agent memory vector collections. Implements three composable DriftDetector
variants (CentroidDrift, PsiDrift, CoherenceDrift) under a shared trait.
All 13 tests pass; benchmark ACCEPTANCE RESULT: PASS (9/9).

Closes nightly research loop for 2026-06-11.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants