Skip to content

ECHO LAYER CANON

ECHO Layer Architecture Canon

Preamble

The ECHO (Epistemically Cached Heuristic Outcomes) Layer is a constitutional knowledge substrate that architecturally eliminates AI hallucinations through multi-sentinel consensus and cryptographic provenance.

Core Principles

  1. Verifiability Over Velocity: Every answer must be verifiable, even at the cost of latency
  2. Epistemic Grounding: Knowledge must be traceable to sources
  3. Consensus Redundancy: No single point of failure in truth assessment
  4. Iterative Refinement: Continuous validation prevents epistemic rot

System Components

Tri-Sentinel Architecture

Primary Sentinel A →
                  ↘
Primary Sentinel B → → Validator → Consensus → GI Scoring → Cache/Escalate
                  ↗
Fallback Sentinel C (DriftGuard)

Ground Integrity (GI) Formula

GI = (Confidence × 0.5) + (Agreement × 0.3) + (SourceQuality × 0.2)

Where:
- Confidence: Average sentinel confidence (0-1)
- Agreement: Sentinels voting for winning answer (0-1)
- SourceQuality: Unique verified sources normalized (0-1)

DriftGuard Protocol

if (GI < 0.85 || !hasRecentSources) {
  activateFallbackSentinels();
  recomputeConsensus();
  if (driftSeverity > 0.15) {
    flagForUrgentHumanReview();
  }
}

Memory Validator

  • Runs every MAX_CACHE_AGE_MS (default: 7 days)
  • Revalidates stale entries through full ECHO review
  • Updates cache or flags for human review
  • Prevents epistemic rot

Database Schema

See infra/db/migrations/20251125_add_echo_layer.sql for full schema.

Consensus Flow

  1. Query Reception: Canonicalize query → Generate embedding → Check cache
  2. Dual-Sentinel: Two primary sentinels answer in parallel
  3. Validation: Validator reviews answers and votes
  4. Consensus: Compute GI score and merge sources
  5. Decision:
  6. GI ≥ 0.85: Cache and approve
  7. GI 0.70-0.84: Human review queue
  8. GI < 0.70: Block and escalate
  9. DriftGuard: Fallback validation if GI is borderline

Security Considerations

  • All cache writes require consensus from ≥3 sentinels
  • Human review queue is encrypted at rest
  • Source URLs undergo domain validation
  • GI scores are cryptographically signed
  • Rate limiting per agent/domain

Scaling Strategy

  • Horizontal: Shard by domain, replicate validators
  • Caching: Redis layer for hot entries
  • Embeddings: Pinecone/Weaviate for vector search
  • Consensus: Async processing for non-critical paths

Observability

DVA.LITE Integration:

interface EchoMetrics {
  giScore: number;
  consensusAgreement: number;
  sourceQuality: number;
  processingTimeMs: number;
  driftScore: number;
  cacheHit: boolean;
}

Governance

Sentinel Selection Criteria

  1. Primary: Must achieve >90% accuracy on TruthBench
  2. Validator: Must detect >95% of adversarial examples
  3. Fallback: Must have independent architecture (different model family)

GI Threshold Calibration

  • Baseline: 0.70 (minimum acceptable truth)
  • Human review: 0.85 (optimal precision/recall)
  • Strict: 0.95 (high-stakes domains)

Freshness Rules

const FRESHNESS_RULES = {
  "breaking-news": 1,    // 1 day
  "medical": 30,         // 30 days
  "legal": 90,           // 90 days
  "scientific": 180,     // 6 months
  "historical": 365      // 1 year
};

Future Enhancements

  1. Federated ECHO: Cross-organizational consensus
  2. Zero-Knowledge Proofs: Privacy-preserving validation
  3. Neurosymbolic Integration: Hybrid reasoning
  4. Active Learning: Sentinel improvement from human reviews
  5. Causal Validation: Counterfactual reasoning checks