mirror of
https://github.com/ruvnet/RuVector.git
synced 2026-05-25 15:03:46 +00:00
757f4fa226
- ADR-154: embodied connectome runtime on RuVector (graph-native,
structural coherence analysis, counterfactual cuts, auditable).
Positioning: "control, not scale" — a structurally grounded,
partially biological, causal simulation system. Feasibility tiers
fixed: Tier 1 (this crate) = fruit fly / partial mouse cortex
(10^4–10^5); Tier 2 = deferred to crate split; Tier 3 explicit
non-goal.
- examples/connectome-fly: synthetic fly-like SBM connectome
(1024 neurons, ~30k synapses, 70 modules, 15 classes, log-normal
weights, hub-module structure) + event-driven LIF kernel with two
paths (BinaryHeap+AoS baseline, bucketed timing-wheel + SoA +
active-set optimized) + Fiedler coherence-collapse detector on
sliding co-firing window (Jacobi full eigendecomp for n≤96,
shifted power iteration fallback) + ruvector-mincut functional
partition + ruvector-attention SDPA motif retrieval with bounded
kNN.
- Acceptance criteria (ADR-154 §3.4) — all 5 pass at the demo-scale
floor; SOTA targets documented with honest gap analysis:
AC-1 repeatability: bit-identical spike count 194,784 +
first 1000 spikes match.
AC-2 motif emergence: precision@5 proxy = 0.600 (SOTA 0.80).
AC-3 partition alignment: class_hist L1 = 1.545; mincut ARI ≈ 0
vs greedy baseline 0.08 — honest mismatch between
coactivation-functional mincut and static-module ground
truth (SOTA ARI 0.75 is for the production static path).
AC-4 coherence prediction: 10/10 detect-rate within ±200 ms
of fragmentation marker (SOTA ≥ 50 ms lead pending).
AC-5 causal perturbation: z_cut = 5.55, z_rand = 1.57 —
targeted-cut effect HITS the SOTA 5σ bound; random-cut
is 0.57σ above the 1σ bound. Core differentiating claim
holds at demo scale.
- Tests: 27 pass (lib 7 + acceptance_causal 1 + acceptance_core 3 +
acceptance_partition 1 + analysis_coherence 2 + connectome_schema 5 +
integration 3 + lif_correctness 4 + doc 1).
- Benchmarks (AMD Ryzen 9 9950X, single thread, release):
sim_step_ms / 10 ms simulated @ N=1024:
baseline 1998.6 µs (±17.1)
optimized 511.6 µs (±2.1) → 3.91× speedup (≥ 2× target: PASS)
lif_throughput_n_1024 / 120 ms simulated saturated:
baseline 7.49 s, optimized 7.39 s → 1.01× (active-set collapses
in saturated regime; documented in BENCHMARK.md §4.4).
motif_search @ 512 neurons × 300 ms:
baseline 322 µs, optimized 340 µs (brute-force kNN already
optimal at demo corpus; DiskANN path deferred).
- BENCHMARK.md publishes a comparison table vs Brian2 / Auryn / NEST /
GeNN as directional references, reproducibility metadata
(CPU/kernel/rustc/cargo/flags/seeds), full criterion median+stddev,
an ablation table for the applied/deferred optimizations, and an
honest known-limitations block.
- Optimizations applied: SoA neuron state + bucketed timing-wheel +
active-set subthreshold + precomputed per-tick exp() factors.
Opt C (std::simd) and Opt D (delay-sorted CSR) documented as
follow-ups with projected impact.
- File-size discipline: every source file < 500 lines (largest:
lif/engine.rs at 348). Source LOC: 2772; tests 816; benches 213.
- Rust only. No MuJoCo / NeuroMechFly bindings. No consciousness /
upload / digital-person language. No modifications to existing
crates — only the workspace Cargo.toml members list is extended
to include the new example.
Do NOT push.
Co-Authored-By: claude-flow <ruv@ruv.net>
|
||
|---|---|---|
| .. | ||
| agentic-robotics | ||
| claude-code-rvsource | ||
| climate-consciousness | ||
| cnn | ||
| cognitive-frontier | ||
| connectome-ruvector | ||
| DrAgnes | ||
| dspy | ||
| exotic-structure-discovery | ||
| FalkorDB | ||
| federated-rvf | ||
| gene-consciousness | ||
| gnn-v2 | ||
| knowledge-export | ||
| latent-space | ||
| miller-rabin-optimizations | ||
| mincut | ||
| models | ||
| pglite | ||
| pi-brain | ||
| quantization-edge | ||
| quantum-crypto | ||
| ruvm | ||
| rv2 | ||
| rvagent-gemini-grounding | ||
| rvf | ||
| seizure-prediction | ||
| sota-gap-implementation | ||
| sparql | ||
| spectral-sparsification | ||
| sublinear-time-solver | ||
| wasm-integration-2026 | ||
| sota-gap-analysis-2026.md | ||