A dated record of the major systems implemented in ruvnet/RuVector, each tied to the first commit that introduced it (hash + author date + link). Two kinds of claim appear, kept separate:
- Provenance (hard fact): the commit hash and author date establish a public, timestamped record of first implementation in this repository. These are verifiable — click any commit.
- Claimed novelty (scoped): where a system appears to be a first-of-its-kind combination or application, the claim is stated with its scope and its priority date (the commit), and known external prior art is named. These are claims anchored to a dated public artifact, not assertions that the underlying concepts (HNSW, LoRA, Merkle logs, hyperbolic embeddings, relational-time physics) were invented here.
Dates are commit author dates (local). Repository genesis: ea3e70aa — Initial commit, 2025-11-19.
Companion documents: ruvnet org-wide prior-art dossier (247 repos across the whole org) · emergent-time overview · witness-chain SOTA survey · ecosystem release notes.
| Date | System | First commit | What it is |
|---|---|---|---|
| 2025-11-19 | Repository genesis | ea3e70aa |
Initial commit |
| 2025-11-26 | ruvector-gnn | 2e4eafea |
Graph neural network crate + compression + WASM/Node bindings |
| 2025-11-30 | Attention mechanisms | e631d4b5 |
ruvector-attention (50+ mechanisms) |
| 2025-12-02 | ruvector-postgres | 286956e7 |
Postgres vector extension + SIMD (#42) |
| 2025-12-03 | SONA self-learning | d249daba |
Self-Optimizing Neural Architecture + RuvLLM |
| 2025-12-23 | Dynamic min-cut | 93ba96e9 |
Subpolynomial-time dynamic minimum-cut system (#74) |
| 2025-12-29 | ReasoningBank / Neural DAG | 85eb5c6e |
Neural Self-Learning DAG + experience store (QuDAG) |
| 2025-12-31 | Intelligence engine | c0b65498 |
Self-learning hooks: trajectory, co-edit, attention |
| 2026-01-15 | Hyperbolic HNSW | 053f4614 |
Poincaré-ball embeddings + HNSW search (#114) |
| 2026-01-20 | Model routing (RuvLTRA) | 96590a1d |
3-tier model routing (#123) |
| 2026-01-27 | AIDefence | 5f9d5fc6 |
Adversarial / prompt-injection protection |
| 2026-02-14 | RVF format | f8870b3c |
RuVector Format — Universal Cognitive Container SDK (#166) |
| 2026-02-25 | Proof-gated graph transformer | 4b79444b |
Graph transformer with 8 verified modules |
| 2026-02-27 | RVF federation | 6990be76 |
Federated transfer learning over RVF |
| 2026-03-03 | pi.ruv.io shared brain | 0c0ef48e |
Shared knowledge-graph MCP server (#234) |
| 2026-03-11 | ruvector-cnn | e743785c |
CNN contrastive learning + SIMD (#252) |
| 2026-03-14 | RuVix microhypervisor | c8803973 |
CLI, kernel shell, PBFT consensus (#261) |
| 2026-03-20 | Spectral sparsifier | 2fd0e743 |
Dynamic spectral graph sparsification |
| 2026-04-04 | RVM hypervisor | a929fde6 |
Coherence-Native Microhypervisor |
| 2026-04-06 | DiskANN/Vamana | 8fbe7686 |
Vamana ANN + PQ, 1.0 recall, 90µs search (#334) |
| 2026-04-23 | RaBitQ quantization | f2dbb6ef |
RaBitQ rotation-based 1-bit quantization (ADR-154) |
| 2026-06-11 | SONA drift gate | 445b43f2 |
Cross-implementation behavioral parity CI gate (#554) |
| 2026-06-12 | Witness v2 | efa3d097 |
Keyed-BLAKE3 chained seals + Merkle + C2SP checkpoint (#558) |
| 2026-06-13 | emergent-time | 44a836d5 |
Calculus of emergent time + Agentic Time primitive (#561) |
Each entry states precisely what is claimed novel, anchored to the dated commit above. Where the building blocks are known external work, that prior art is named — the claim is the specific synthesis, not the parts.
Priority date: 2026-02-14 (f8870b3c) · ADR-029 / ADR-030 · crates/rvf/
Claim: a self-describing on-disk vector format that unifies, in one portable container, the raw vectors, the ANN index (HNSW), optional quantization, and cognitive/provenance metadata — designed so the serialized layout is the search layout and the artifact carries its own witness/integrity record.
Prior art: FAISS index files, usearch serialized indexes, Lance/Arrow columnar vector storage. Scoped novelty: the "cognitive container" framing (a vector artifact as a self-contained, provenance-bearing cognitive unit) and the unified index+quant+witness packaging.
Priority date: 2026-04-04 (a929fde6) · crates/rvm/
Claim: a no_std microhypervisor whose partitioning/scheduling is driven by a coherence engine (graph min-cut over agent state) and whose every privileged mutation is proof-gated and recorded in a tamper-evident witness chain ("no witness, no mutation").
Prior art: seL4 (verified microkernel), Microsoft CCF (TEE ledger), Firecracker (microVM). Scoped novelty: coherence-native scheduling (mincut-driven partition decisions) combined with proof-gated mutation and a syscall-path witness log as one architecture.
Priority date: 2026-06-13 (44a836d5) · ADR-251 · crates/emergent-time/ · crates.io · npm
Claim: a software primitive that defines an AI agent's internal time as arc length through its own state manifold — grounded in relational-time physics (Wheeler–DeWitt, Page–Wootters, Connes–Rovelli thermal time) — with explainable per-channel ticks, an Agentic Time Index, and a health classifier.
Prior art: the relational-time formalisms (1967–1994); ADWIN / concept-drift detection. Scoped novelty: the unified physics-grounded framing as a runtime control/diagnostic primitive for agents.
Honest caveat (prove-not-hype): on real recorded agent traces, with pre-registered thresholds and an error-blinded clock, this primitive does not beat a fair change-point baseline on early-warning lead (0 win / 1 tie / 1 loss fixed-window; 0/0/2 with an adaptive detector). Its demonstrated value is the explainable diagnostic layer, not a lead win. Stated in the crate, its tests, and ADR-251.
Priority date: 2025-12-03 (d249daba) · ADR-074 · crates/sona/
Claim: a routing/recall self-optimizer combining a MicroLoRA "instant loop," EWC++ catastrophic-forgetting protection, and a ReasoningBank experience store, adapting from a single feedback signal at sub-millisecond cost.
Prior art: LoRA, EWC, experience replay. Scoped novelty: the specific sub-ms instant-loop composition for online LLM-router self-optimization.
Priority date: 2026-06-11 (445b43f2) · scripts/sona-drift/ (#554)
Claim: a CI gate that detects behavioral drift between parallel implementations of the same learning algorithm by storing each implementation's behavioral fingerprint as vectors in the project's own format (RVF / reference.rvf) and failing the build on vector-distance drift — plus a static "tripwire" guarding the falsifiability tests themselves.
Prior art: golden-file/snapshot testing; metamorphic testing. Scoped novelty: using a vector database's own format to store behavioral fingerprints and gate cross-implementation semantic drift in CI (dogfooding the format as test infrastructure). Appears to be genuinely first-of-its-kind.
Priority date: 2026-06-12 (efa3d097) · ADR-134 · crates/rvm/crates/rvm-witness/
Claim: a witness log with 128-bit keyed-BLAKE3 chain links, Merkle segment sealing, publicly verifiable cross-segment binding, forward-secure key ratcheting, and C2SP tlog-checkpoint export — running on a no_std kernel syscall path at ~112 ns/append.
Prior art: Certificate Transparency (RFC 6962/9162), Sigstore Rekor, Trillian/Sunlight tile logs, Microsoft CCF receipts. Scoped novelty: bringing transparency-log-grade, externally-anchorable provenance to a sub-microsecond kernel audit path, with measured numbers. (SOTA survey.)
Priority date: 2026-01-15 (053f4614) · crates/ruvector-hyperbolic-hnsw/ (#114)
Claim: HNSW graph search operating natively in the Poincaré ball (hyperbolic distance), for embeddings of hierarchical/tree-structured data.
Prior art: hyperbolic embeddings (Nickel–Kiela 2017); HNSW (Malkov–Yashunin 2016). Scoped novelty: integrating Poincaré-ball geometry with HNSW graph traversal as a usable ANN index.
Priority date: 2026-01-27 (5f9d5fc6) · npm/packages/ruvbot/src/security/
Claim: a runtime defense layer that scans agent inputs/outputs for prompt-injection / jailbreak / manipulation, with adaptive mitigation.
Prior art: prompt-injection detection research; guardrail libraries. Scoped novelty: an integrated, agent-runtime manipulation-defense layer wired into the swarm/witness stack.
Priority date: 2025-12-23 (93ba96e9) · crates/ruvector-mincut/ (#74)
Claim: a production Rust implementation of subpolynomial-time dynamic minimum cut, later wired into RVM's coherence engine.
Prior art: recent algorithmic theory (STOC/FOCS-era). Scoped novelty: implementation-first — a usable engineered version of the theory.
These are real first-implementations in the repo; the underlying ideas are established external work, so they are listed for the record without a global-first claim.
- ruvector-postgres — Postgres vector extension with SIMD + 53 SQL functions. 2025-12-02 (
286956e7). Prior art: pgvector. - ReasoningBank / Neural Self-Learning DAG — trajectory → verdict → distill experience store. 2025-12-29 (
85eb5c6e); an earlier learning module appears inruvector-postgres(2025-12-02). - Intelligence engine / hooks pipeline — RETRIEVE → JUDGE → DISTILL → CONSOLIDATE self-learning hooks. 2025-12-31 (
c0b65498). - Model routing (RuvLTRA) — 3-tier model routing. 2026-01-20 (
96590a1d). - RVF federation — federated transfer learning. 2026-02-27 (
6990be76). - pi.ruv.io shared brain — collective knowledge-graph MCP server with differential privacy. 2026-03-03 (
0c0ef48e). - DiskANN/Vamana, RaBitQ, ruvector-cnn, ruvector-gnn, graph-transformer, attention, sparsifier — see timeline; all implementations of named external methods.
crates.io: rvf-runtime 0.3.0 · rvf-index 0.2.0 · rvf-quant 0.2.0 · rvf-types 0.2.1 · emergent-time 2.2.4
npm: ruvector 0.2.30 · @ruvector/rvf 0.2.2 · @ruvector/rvf-wasm 0.1.7 · @ruvector/sona 0.1.7 · @ruvector/ruvllm 2.5.6 · @ruvector/emergent-time 0.1.0
Recent hardening (2026-06-11/13) landed measured wins: RVF queries wired to HNSW (~14× faster, recall@10 0.968), a contiguous vector slab (brute scan ~6.4× faster), RaBitQ two-stage search (recall@10 0.972 at 32× compression), the witness v2 work, and the emergent-time crate (~53× faster Schrödinger path) — each with tests and benchmarks rather than asserted numbers.
- AgentDB is NOT a RuVector crate. The vector-memory database referenced throughout (
mcp__claude-flow__agentdb_*tools, the skills) lives in the separate claude-flow / agentic-flow ecosystem; RuVector integrates with it but does not contain it. It is excluded from these claims to keep provenance accurate. - Multi-agent swarm / hive-mind orchestration is also NOT in this repository — same external
@claude-flow/cliecosystem. The only "swarm" code in RuVector's history is application-level (e.g.ruvector-roboticscoordination examples). - "First commit" = earliest commit touching the path on the default branch; squash-merges may collapse earlier development into a single dated commit, so these dates are upper bounds on first public appearance.
- Every commit link above resolves to a real, verified commit. Novelty claims are scoped and dated; constituent concepts are credited to their external originators.
Generated from git history of ruvnet/RuVector. Provenance is fact; novelty claims are scoped to the stated combination/application and anchored to a public dated commit.
