Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Train SONA + MicroLoRA neural patterns from successful task completions; runs the DISTILL + CONSOLIDATE phases of the 4-step pipeline
.claude/skills/ruvnet-neural-train/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-11 | ✗→✓ | ▲ Improved | -16% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 67% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -11% | 0% |
Train and consolidate neural patterns. Implements the DISTILL and CONSOLIDATE phases of the 4-step intelligence pipeline.
mcp__plugin_ruflo-core_ruflo__neural_status.mcp__plugin_ruflo-core_ruflo__hooks_intelligence_trajectory-start with the task context.mcp__plugin_ruflo-core_ruflo__hooks_intelligence_trajectory-step.mcp__plugin_ruflo-core_ruflo__hooks_intelligence_trajectory-end with verdict: pass|fail|partial.mcp__plugin_ruflo-core_ruflo__hooks_intelligence_learn.mcp__plugin_ruflo-core_ruflo__neural_train with --pattern-type coordination --epochs 10.mcp__plugin_ruflo-core_ruflo__hooks_intelligence_pattern-store.mcp__plugin_ruflo-core_ruflo__neural_patterns to confirm.For real-time micro-adaptation:
bashmcp tool call ruvllm_sona_create --json -- '{"domain": "coding"}' mcp tool call ruvllm_sona_adapt --json -- '{"feedback": {"score": 0.9, "trajectory": "..."}}'
When you have ≥3 distinct domains, create a MicroLoRA adapter per domain rather than overloading SONA:
bash# Create the adapter mcp tool call ruvllm_microlora_create --json -- '{"domain": "frontend"}' # Adapt with feedback mcp tool call ruvllm_microlora_adapt --json -- '{"adapter": "frontend", "feedback": {...}}' # CONSOLIDATE phase: apply EWC++ on weight deltas to prevent catastrophic forgetting mcp tool call ruvllm_microlora_adapt --json -- '{"adapter": "frontend", "consolidate": true}'
The --consolidate flag is the EWC++ trigger. Without it, fresh training overwrites older domains.
After every ~10 trajectory completions, run a full consolidation pass:
bashmcp tool call agentdb_consolidate --json mcp tool call neural_compress --json # storage efficiency
This folds patterns into long-term storage under EWC++ semantics.
If the system has no learned patterns yet:
bashmcp tool call hooks_pretrain --json -- '{"modelType": "moe", "epochs": 10}' mcp tool call hooks_build-agents --json -- '{"agentTypes": "coder,tester"}'
hooks_pretrain writes to the patterns (plural) namespace — distinct from the pattern (singular) ReasoningBank target. See ruflo-agentdb ADR-0001 for the namespace convention.
To wipe intelligence state (e.g., for benchmarking):
bashmcp tool call hooks_intelligence-reset --json
bashnpx @claude-flow/cli@latest neural train --pattern-type coordination --epochs 10 npx @claude-flow/cli@latest neural patterns --list npx @claude-flow/cli@latest neural status npx @claude-flow/cli@latest neural compress npx @claude-flow/cli@latest hooks pretrain --model-type moe --epochs 10 npx @claude-flow/cli@latest hooks build-agents --agent-types coder,tester
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | fail→pass | 16,398 | 1,626 | -90% | 1 | 1 | 0% | 1,525 | 1,288 | -16% | 0 | 0 | — |
case-01 | fail→fail | 10,551 | 4,719 | -55% | 1 | 1 | 0% | 1,872 | 1,560 | -17% | 0 | 0 | — |
case-02 | fail→fail | 5,596 | 5,146 | -8% | 1 | 1 | 0% | 805 | 1,559 | +94% | 0 | 0 | — |
case-03 | fail→pass | 8,710 | 6,314 | -28% | 1 | 1 | 0% | 1,384 | 2,311 | +67% | 0 | 0 | — |
case-04 | fail→pass | 10,714 | 3,405 | -68% | 1 | 1 | 0% | 950 | 1,672 | +76% | 0 | 0 | — |
case-05 | fail→pass | 18,005 | 4,426 | -75% | 1 | 1 | 0% | 3,689 | 1,895 | -49% | 0 | 0 | — |
case-06 | fail→pass | 9,180 | 2,359 | -74% | 1 | 1 | 0% | 1,640 | 1,463 | -11% | 0 | 0 | — |
case-07 | fail→pass | 13,335 | 6,462 | -52% | 1 | 1 | 0% | 2,198 | 2,505 | +14% | 0 | 0 | — |
case-08 | fail→pass | 16,424 | 1,529 | -91% | 1 | 1 | 0% | 2,813 | 1,259 | -55% | 0 | 0 | — |
case-09 | fail→pass | 8,093 | 1,729 | -79% | 1 | 1 | 0% | 1,379 | 1,298 | -6% | 0 | 0 | — |
case-10 | fail→pass | 6,264 | 1,298 | -79% | 1 | 1 | 0% | 1,099 | 1,241 | +13% | 0 | 0 | — |
case-12 | fail→pass | 8,292 | 1,713 | -79% | 1 | 1 | 0% | 1,353 | 1,246 | -8% | 0 | 0 | — |
case-13 | fail→pass | 8,134 | 1,596 | -80% | 1 | 1 | 0% | 1,386 | 1,272 | -8% | 0 | 0 | — |
case-14 | fail→pass | 10,188 | 2,023 | -80% | 1 | 1 | 0% | 1,686 | 1,315 | -22% | 0 | 0 | — |
case-15 | fail→pass | 10,000 | 1,899 | -81% | 1 | 1 | 0% | 1,613 | 1,259 | -22% | 0 | 0 | — |
case-16 | pass→pass | 9,231 | 1,489 | -84% | 1 | 1 | 0% | 1,663 | 1,218 | -27% | 0 | 0 | — |
case-17 | fail→pass | 4,451 | 1,433 | -68% | 1 | 1 | 0% | 803 | 1,251 | +56% | 0 | 0 | — |
case-18 | fail→pass | 8,334 | 1,772 | -79% | 1 | 1 | 0% | 1,373 | 1,304 | -5% | 0 | 0 | — |
case-19 | fail→pass | 6,896 | 1,364 | -80% | 1 | 1 | 0% | 1,047 | 1,181 | +13% | 0 | 0 | — |
case-20 | fail→fail | 17,821 | 13,966 | -22% | 1 | 1 | 0% | 3,776 | 3,876 | +3% | 0 | 0 | — |
case-21 | pass→pass | 11,901 | 9,855 | -17% | 1 | 1 | 0% | 2,350 | 2,986 | +27% | 0 | 0 | — |
case-22 | pass→pass | 14,428 | 14,634 | +1% | 1 | 1 | 0% | 3,062 | 4,240 | +38% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +73 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.