Starstruk Logo MarkStarstruk

v3.10.7

ruvnet/ruflov3.10.7May 29, 2026by ruvnet

AI Summary

Intelligence self-learning audit with critical fixes and honest performance numbers. Fixes negative-reward inversion and removes fabricated metrics.

Key Highlights

  • Critical: negative-reward inversion fixed (-1.0 now correctly parsed)
  • Removed fabricated Flash Attention speedup metric (was RNG-based)
  • Embedding observability: generateEmbedding returns backend: onnx|mock
  • MCP learning: trajectory-end no longer feeds synthetic gradient to EWC
  • HNSW optimization: 3.2–4.7x improvement at N=5000, 1.89x at N=20000

New Features

  • Negative numeric flag value parsing
  • Embedding backend visibility
  • Real EWC distill/consolidate cycle
  • HNSW optimization with proper configuration

Full Release Notes

## 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-coordinator` copies; now an honest "unmeasured" sentinel.
- **Embedding observability** — `generateEmbedding` returns `backend: onnx｜mock`, surfaced in `memory_bridge_status`/`import` so a mock (hash-fallback) embedding is never mislabeled as the real ONNX model.
- **MCP learning** — `trajectory-end` no longer feeds EWC a synthetic gradient; `hooks_intelligence_learn` runs 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](https://github.com/ruvnet/ruflo)