research: mincut-partitioned agent-memory consolidation (nightly 2026-08-17) — hypothesis rejected, two ruvector-mincut defects found

[ruvnet]

Summary

Nightly research run testing whether partitioning the agent-memory similarity graph before applying a retention budget protects minority topics from being evicted in full by ruvector-agent-memory's global top-score CoherencePolicy (nightly 2026-06-14).

• Hypothesis: REJECTED by direct measurement, at every scale tested (n=4000, n=500, and a bounded floor_min sweep). Worst-cluster recall gain never reached the pre-declared 15pp threshold — best result was 0.00pp at n=4000, 8.00pp at n=500.
• The mechanism is not worthless: overall recall improved +6.8pp and 4 of 6 clusters individually gained 15–23pp recall each at n=4000. The specific cluster the hypothesis exists to protect stayed merged with the majority cluster by the partitioner, so it received no protection — a bounded sweep of the retention floor confirmed this is a partitioning shortfall, not a retention-budget shortfall (worst-cluster recall was flat across every floor value tested).
• Along the way, found and documented two independent, reproducible defects in ruvector-mincut: (1) DynamicMinCut::partition() returns vertex splits inconsistent with its own min_cut_value(), nondeterministically; (2) GraphPartitioner/RuVectorGraphAnalyzer drops vertices, fabricates vertex ids for non-contiguous id spaces, and is severely slow (8.4s at n=500, did not finish in 5m42s at n=4000). Both are worked around (not fixed upstream) with a from-scratch, tested weighted Stoer–Wagner implementation (mincut_exact.rs) used as the authoritative partition source; ruvector_mincut's min-cut value is still cross-checked and was never observed wrong, only its partition materialization.

Architecture

• New crate crates/ruvector-partition-memory (experimental, not wired into any production path).
• Depends on ruvector-agent-memory (wired into the workspace members list — it predated workspace membership and was otherwise unbuildable) for the baseline CoherencePolicy scorer, reused directly rather than re-implemented.
• Depends on ruvector-mincut for GraphPartitioner (candidate A) and as a cross-check source for mincut_exact.rs's own Stoer–Wagner (candidate B).
• witness.rs: SHA-256 hash-chain over partition split decisions, following the precedent set by ruvector-retrieval-receipt (nightly 2026-08-13).

Files changed

• crates/ruvector-partition-memory/ — new crate: corpus.rs, graph.rs, mincut_exact.rs, partition.rs, retention.rs, metrics.rs, witness.rs, search.rs, main.rs (benchmark binary), examples/calibrate.rs, examples/darwin_sweep.rs.
• Cargo.toml / Cargo.lock — add ruvector-partition-memory and (previously orphaned) ruvector-agent-memory to workspace members.
• docs/adr/ADR-305-mincut-partitioned-memory-consolidation.md
• docs/research/nightly/2026-08-17_mincut-partitioned-memory-consolidation/ — README, gist, and raw evidence/*.txt benchmark output.

Benchmark command

cargo run --release -p ruvector-partition-memory --bin benchmark -- 4000 3 0.35 10 600
cargo run --release -p ruvector-partition-memory --bin benchmark -- 500 3 0.35 10 600
cargo run --release -p ruvector-partition-memory --example darwin_sweep

Real benchmark results (n=4000, accepted run)

variant          overall_recall  worst_cluster_recall  coverage
GlobalTopScore   0.4193          0.1520                 1.000
MincutAdaptive   0.4873          0.1520                 1.000

per_cluster_recall GlobalTopScore = [0.996, 0.152, 0.216, 0.316, 0.396, 0.440]
per_cluster_recall MincutAdaptive = [0.792, 0.380, 0.504, 0.152, 0.556, 0.540]

worst_cluster_gain_pp = -0.00  (threshold: +15.00)
ACCEPTANCE_RESULT: REJECT

Full raw output for this run, the n=500 run, the floor_min Darwin sweep, and the corpus-noise calibration pass are in docs/research/nightly/2026-08-17_mincut-partitioned-memory-consolidation/evidence/.

Acceptance result

REJECT. See ADR-305 for the full rejection criteria, evidence, and a proposed follow-up hypothesis (a per-branch/size-weighted stopping criterion) — not implemented here, since changing the criterion after seeing this run's results would defeat the point of a pre-declared threshold.

Darwin result

Bounded sweep (generations=1, candidates=4) over floor_min ∈ {1,3,8,15}, partition held fixed. Winner floor_min=15 on a composite fitness metric (0.5·worst + 0.3·overall + 0.2·correctness), but worst_cluster_recall itself did not improve — confirming the primary hypothesis's shortfall is structural, not a retention-budget tuning problem. Not promoted to any default; parent retained.

Flywheel result

No ruvector harness flywheel CLI was found in this environment (verified via npx ruvector harness doctor/flywheel --help, no output/executable resolved) — evidence is instead preserved as plain files under evidence/, referenced from the ADR and README, rather than fabricated through a nonexistent interface.

Security review

No new attack surface — standalone research crate over synthetic data, not wired into any request-serving path. witness.rs is a tamper-evidence mechanism for auditing a partition decision after the fact, not an access-control or correctness proof; this is stated explicitly in the ADR's Security section to avoid overclaiming.

Main limitations

• Single run per configuration (no repeated-trial variance reporting).
• One synthetic corpus/seed family; not validated against a real agent-memory trace.
• The two ruvector-mincut defects are documented with minimal repros but not filed upstream from within this run — that needs the owning maintainer's independent verification.
• No WASM/edge measurement was taken (analysis-only, per the ADR's mandatory-but-optional-implementation guidance).

Production recommendation

Do not promote. Hypothesis rejected by direct measurement. mincut_exact.rs's correct, tested Stoer–Wagner implementation is a reusable asset independent of this ADR's outcome for any future RuVector graph-partitioning work.

Research documents

• ADR: docs/adr/ADR-305-mincut-partitioned-memory-consolidation.md
• Nightly research README: docs/research/nightly/2026-08-17_mincut-partitioned-memory-consolidation/README.md
• Gist: docs/research/nightly/2026-08-17_mincut-partitioned-memory-consolidation/gist.md

---

Generated by Claude Code