research(nightly): distance-adaptive beam search — REJECT with evidence

[ruvnet]

Hypothesis

ADR-303 (ruvector-entropy-ann) rejected Shannon-entropy-based adaptive beam-width control for ANN
graph traversal, but its prior-art table cited an untested alternative: the (1+γ)·d_k
relative-distance stopping rule from Distance Adaptive Beam Search for Provably Accurate
Graph-Based Nearest Neighbor Search (arXiv:2505.15636, NeurIPS 2025). This nightly implements and
measures that alternative, on the same dataset/graph construction ADR-303 used, subject to the same
matched-budget-control discipline that caught the entropy signal's flaw.

Pre-registered acceptance (fixed before any run):

1. Per-query cost must vary more on hard queries than easy queries (ratio ≥ 1.15) — the direct test
   for "does it actually adapt."
2. Recall within 3 points of a high-recall FixedEf(100) reference on every query set.
3. On hard queries, beat a FixedEf baseline calibrated (on a disjoint query set) to the same
   average cost, by ≥ 2 recall points.

Architecture

• crates/ruvector-dab-search: FixedEf (baseline), AdaptiveGamma uncapped (candidate A),
  AdaptiveGamma capped (candidate B, production safety bound).
• Same flat k-NN graph construction as ADR-303, but with a corrected entry-routing scheme
  (entry_seeds) — a real bug was found and fixed during development: a single fixed entry point
  cannot reach graph components/clusters it isn't connected to (measured 19% recall before the fix).
• Zero external dependencies, deterministic fixed-seed dataset generation throughout.

Benchmark

cargo test --release -p ruvector-dab-search      # 17/17 pass
cargo run --release -p ruvector-dab-search --bin benchmark

Real, reproducible numbers (two independent runs matched to printed precision):

Query set	FixedEf(50) recall	FixedEf(100) recall	Adaptive(γ=0.5) recall	Adaptive mean dist_comp
easy	0.811	0.846	0.903	346.7 (sd 153.6)
hard	0.706	0.722	0.756	317.3 (sd 130.4)
mixed	0.635	0.663	0.678	291.5 (sd 95.7)

• Test 1 (adapts per query): FAIL. hard/easy cost ratio = 0.915 (needed ≥1.15) — the rule does
  vary substantially per query (unlike ADR-303's constant ~124), but in the wrong direction: it
  tracks local point density, not task difficulty.
• Test 2 (recall floor): PASS on all three query sets.
• Test 3 (matched-budget crux test): FAIL, narrowly. +0.014 recall advantage vs required +0.02.
• Headline (cost at matched recall, the source paper's own metric): -6.6% — needs more work
  than a plain fixed budget for equal recall, the opposite of the paper's reported 10–50% reduction
  (measured on real navigable HNSW graphs over real embeddings — this PoC's flat exact-k-NN graph is
  not proven navigable, which is the leading candidate explanation; see ADR-340's Open Questions).

Acceptance Result

REJECT (pre-registered logic: Test 1 or Test 2 failing → REJECT; only Test 3 failing →
INCONCLUSIVE. Test 1 failed outright, so REJECT applies.)

Darwin / Flywheel

No Darwin evolution phase was run (Darwin/MetaHarness CLI tooling was probed —
npx metaharness --help resolves to the harness-scaffolding generator, not a research-orchestration
API for this repo; npx ruvector harness doctor does not resolve in this environment) — this
capability was verified absent rather than assumed. The negative result and its evidence are
retained directly in ADR-340 and the research README as this run's Flywheel-equivalent record,
including the specific mechanism (density vs. difficulty confound) so a future attempt doesn't
rediscover it blindly.

Security Review

No security-relevant surface: benchmark-only crate, no I/O, no network access, no production
integration path, no secrets or external data.

Main Limitations

• Single synthetic dataset (N=2,000, dim=16, 10 clusters); no real embedding benchmark this run.
• Flat exact-k-NN graph, not an incrementally-built HNSW/Vamana graph — navigability (the source
  theorem's precondition) is not verified to hold.
• Single hardware configuration; no SIMD/multi-thread path measured.

Production Recommendation

Do not adopt in this PoC form. If revisited: test on a real incrementally-built HNSW graph and real
embedding dataset before drawing conclusions about the method itself (see ADR-340's Open Questions
and "If this is ever revisited" sections).

Files

• crates/ruvector-dab-search/ — implementation, tests (17, all passing), benchmark binary
• docs/adr/ADR-340-distance-adaptive-beam-search.md — ADR
• docs/research/nightly/2026-08-25-distance-adaptive-beam-ann/README.md — full research report
• docs/research/nightly/2026-08-25-distance-adaptive-beam-ann/gist.md — standalone technical article
• Cargo.toml / Cargo.lock — workspace member registration

---

Generated by Claude Code