Skip to content

v3.24.0 — Self-Learning Flywheel

Choose a tag to compare

View all tags
@ruvnet ruvnet released this 05 Jul 15:34
v3.24.0
2ed7f7f

ruflo 3.24.0 — The Self-Learning Flywheel

ruflo can now improve one of its own operating policies over time and prove each improvement is real — not marketing. Merges #2572 (ADR-176 self-optimizing flywheel + ADR-177 signed config propagation).

📖 Full write-up (plain language + technical + usage + upgrade notes): https://gist.github.com/ruvnet/f8e2851fd307df5d5de7b5c70c37fa0c

What's new

  • Verified retrieval improvement, auto-applied to every install. A signed config champion (Ed25519 + RVFA) is adopted on startup, fail-closed on authenticity and suitability. Better retrieval defaults out of the box, +0.0738 nDCG@3 over the previously-tuned baseline. No re-init needed.
  • Self-optimizing flywheel (opt-in, $0 default). The background daemon compounds verified retrieval-policy improvements: each generation reads the persisted champion as baseline, gates a candidate on a frozen held-out with a significance test (accept/v1+sig) + human-relevance guard + a separate canary, and on promotion advances the champion so the next tick builds on it. Winners accumulate into a signed, independently-replayable lineage back to an immutable root (git-for-operating-policies).
  • Shadow-first / no auto-serve + drift canary. Promoted champions serve only after a one-generation shadow delay; a canary re-scores on the evolving store each tick and auto-rolls-back regressions.
  • Meta-learning. The optimizer biases its search toward policy axes with measured historical payoff.
  • Proof, not assertion. Receipt bundles replay independently without trusting our logs; a CI guard keeps the shipped evidence valid on every PR.

Demonstrated live: two real, significant, compounding promotions (self-retrieval RR 0.496 → 0.758 → 0.847), human relevance preserved, zero human intervention.

Honest scope: the flywheel's compounding gains are on a self-supervised retrieval benchmark, gated so human-labeled relevance does not regress — not a claim that human relevance improved generation-over-generation (held flat by design). The auto-applied one-shot champion was tuned on human-labeled relevance.

Upgrade

npx ruflo@latest        # or npx ruflo@3.24.0

Backwards-compatible / additive · signing keys unchanged · the flywheel is off unless RUFLO_HARNESS_LOOP=1.

Packages: @claude-flow/cli@3.24.0 · claude-flow@3.24.0 · ruflo@3.24.0

🤖 Generated with RuFlo

Assets 2
Loading