ADR-075 — Unified Learning Stats Aggregator (#2245 follow-up)

Status: Accepted — Implemented in ruflo 3.10.15 Date: 2026-05-30 Tracking: #2245 — self-learning subsystem audit Supersedes: the "future-rounds" item in ADR-074 §"Deliberately NOT in this round" titled "Unify the 4 stat sources"

Context

ADR-074 (#2245 round 1) wired the three broken self-learning surfaces (hooks_task-completed, signalsProcessed, pretrain) and made every handler return an honest learningPath/sources.stores/note so callers see which path ran. It left one item to a future round:

Unify the 4 stat sources — designate globalStats as authoritative, derive memory_bridge_status, hooks metrics, and neural_patterns count from it. Multi-PR design work.

When we tried that, we found the premise was wrong: the four sources genuinely measure different things. They look like duplicates of one number because they all answer "did learning happen?" — but each store is the authoritative record for a different layer:

Designating one as authoritative would either lose information or force lossy back-fill. The right move is to aggregate the view, not the store.

Decision

Add a read-only aggregator that calls all four stores and returns them in one shape, with each sub-view explicitly naming its source. No new store. No migration. No data motion. One honest answer.

Public surface

getUnifiedLearningStats() in intelligence.ts returns:

interface UnifiedLearningStats {
  global:        { patternsLearned, trajectoriesRecorded, signalsProcessed, lastAdaptation, source };
  sona:          { trajectoriesTotal, patternsLearned, reasoningBankSize, avgAdaptationTimeMs, source, available };
  memoryBridge:  { totalEntries, perNamespace, source, reachable };
  neuralPatterns:{ patternCount, byType, modelCount, source };
  consistency:   { sonaTracksGlobal, sonaTracksGlobalDelta, notes };
  generatedAt:   string;
}

MCP surface

hooks_intelligence_unified-stats exposes the aggregator to MCP clients. Description names every sub-view's store and the kind of question each answers, so an LLM agent can choose between this (one coherent dashboard) and the original narrow aggregators (one specific slice).

Helpers that fell out

• getMemoryBridgeStats({ namespaces? }) in memory-bridge.ts — read-only top-level + per-namespace entry counts. Best-effort: returns reachable:false instead of throwing when the bridge isn't loadable.
• getNeuralStoreStats() in neural-tools.ts — top-level + per-type pattern count from the neural store file.

Both are exported so anything wanting a coherent learning-state snapshot can build it from the same primitives.

The consistency block

Instead of enforcing an invariant between stores (which would require schema-level work in the data path), we describe the relationships and flag drift:

• sona.trajectoriesTotal is expected to track globalStats.trajectoriesRecorded within ±1 between a process restart and the next clearIntelligence(). Drift outside ±2 is a consistency.notes entry.
• When globalStats.patternsLearned > 0 but neuralPatterns.patternCount === 0, the consistency block reports it explicitly — that's the pretrain-vs-trajectory-vs-neural-store gap the reporter caught in #2245.
• When the memory-bridge is unreachable, that's named in consistency.notes so bridge-dependent counters showing 0 aren't misread as "no activity."

This is the same shape as the per-handler learningPath/note design from ADR-074: report what you did and what you couldn't do, never silently lie.

Deliberately NOT in this ADR

• Schema-level unification (one store, others derived). Wrong design — see Context.
• Persisting sonaCoordinator stats across process restarts. Out of scope; tracked under #2245's "background consolidation worker" item.
• Wiring post-edit / post-command to feed the trajectory pipeline. Tracked as round B of the post-ADR-074 work.
• Structured Distillation of trajectory content (arXiv:2603.13017, #2241). Tracked as round C.

Consequences

• One call answers "did learning happen?" across all 4 layers, with each layer's source path named.
• The consistency.notes array catches the specific drifts the reporter found (the "four contradictory sources" complaint) — they're now surfaced as flags on the unified call rather than as silent disagreement between separate dashboards.
• The original aggregators stay where they are; nothing depends on them being merged. Code paths that need exactly the SONA slice still call hooks_intelligence_stats. Code paths that need the bridge slice still call memory_bridge_status. The unified call exists alongside them.
• No data migration. No back-compat shim. New surface, new MCP tool, additive.

Verification

• __tests__/unified-stats-2245.test.ts — 7 tests:
  • Shape: all four sub-views present with source fields ✅
  • Driving each path moves the right counter:
    • recordSignalProcessed → global.signalsProcessed up ✅
    • storeNeuralPatterns → neuralPatterns.patternCount up ✅
    • hooks_task-completed {trainPatterns:true} → global.trajectoriesRecorded up ✅
  • Consistency block flags drift instead of staying silent ✅
  • MCP tool registered + returns the unified shape ✅
• scripts/benchmark-self-learning.mjs §F — proof artifact in docs/benchmarks/runs/self-learning-latest.json.
• Build clean (tsc -b); full CLI suite green (modulo the 3 pre-existing flakes documented in ADR-074).

Reproduce

git clone https://github.com/ruvnet/ruflo && cd ruflo
npm install && ( cd v3/@claude-flow/cli && npx tsc -b )

# Cross-store consistency tests
( cd v3/@claude-flow/cli && npx vitest run __tests__/unified-stats-2245.test.ts )

# 6-section benchmark including §F unified-stats
node v3/@claude-flow/cli/scripts/benchmark-self-learning.mjs