This release includes breaking changes for platform teams planning a safe upgrade.
✓ No known CVEs patched in this version
Topics
+14 more
Summary
AI summaryUpdates Other fixes, hash-fallback, and https://github.com/ruvnet/ruflo across a mixed release.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Low |
Performance tables now show measured values from benchmark-intelligence.mjs Performance tables now show measured values from benchmark-intelligence.mjs Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Performance | Medium |
HNSW optimization yields 3.2–4.7× speedup for N=5000 and 1.89× for N=20000 HNSW optimization yields 3.2–4.7× speedup for N=5000 and 1.89× for N=20000 Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Bugfix | Critical |
Negative reward flag now correctly parses -1.0 instead of +1.0 Negative reward flag now correctly parses -1.0 instead of +1.0 Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Bugfix | Medium |
Removed fabricated Flash Attention speedup metric; reports "unmeasured" sentinel Removed fabricated Flash Attention speedup metric; reports "unmeasured" sentinel Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Bugfix | Medium |
Embedding observability correctly labels backend as onnx or mock Embedding observability correctly labels backend as onnx or mock Source: llm_adapter@2026-05-29 Confidence: high |
— |
| Bugfix | Medium |
MCP learning now runs real distill/consolidate cycle instead of synthetic gradient MCP learning now runs real distill/consolidate cycle instead of synthetic gradient Source: llm_adapter@2026-05-29 Confidence: high |
— |
Full changelog
Ruflo v3.10.7 — intelligence self-learning audit, hardening fixes & honest performance numbers
A full empirical audit of the self-learning/intelligence system, the prioritized fixes it surfaced, and a rewrite of all performance claims to measured values. Audit + reusable benchmark harness included.
🔴 Critical fix — negative-reward inversion (follow-up to #2222)
route feedback -r -1.0 (and --reward -1.0) was parsed as +1.00 — the shared CLI flag parser dropped any --prefixed value, so giving negative feedback actively reinforced the bad agent. Fixed in parser.ts (negative numeric literals are now accepted as flag values); all three syntaxes yield −1.0. Verified in the published artifact.
Other fixes
- Removed a fabricated metric — Flash Attention "speedup" was reported from a runtime RNG in both
attention-coordinatorcopies; now an honest "unmeasured" sentinel. - Embedding observability —
generateEmbeddingreturnsbackend: onnx｜mock, surfaced inmemory_bridge_status/importso a mock (hash-fallback) embedding is never mislabeled as the real ONNX model. - MCP learning —
trajectory-endno longer feeds EWC a synthetic gradient;hooks_intelligence_learnruns a real distill/consolidate cycle.
HNSW optimization (genuine, measured)
Root cause: HNSW was never actually running — the ruvector adapter passed no storagePath, the native DB's file lock was held by a daemon, and a silent catch{} degraded to brute force. Fixed (unique storagePath, hnswConfig {m:32, efConstruction:200}, visible fallback warning). Same-harness before→after: N=5000 0.92×→3.2–4.7×, N=20000 0.95×→1.89× (recall@10 0.88–0.99).
Honest performance numbers
README + CLAUDE.md perf tables now show measured values from the new scripts/benchmark-intelligence.mjs:
- Int8 quantization 3.84× (reconstruction cosine 0.99999) · RaBitQ 32× memory · SONA adapt 0.0043 ms · MoE gate converges (0.13→0.88)
- HNSW "150×–12,500×" and Flash "2.49–7.47×" marked NOT reproduced / unverified (no benchmark supports them)
Full audit: docs/reviews/intelligence-system-audit-2026-05-29.md. All three packages published at 3.10.7 (latest/alpha/v3alpha in lockstep).
🤖 Generated with RuFlo
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
Related context
Related tools
Beta — feedback welcome: [email protected]