Skip to content

Beyond Algorithmic Trust (Draft)

Context (Oct 2025)
PoGQ + reputation + robust aggregation covers the algorithmic surface, but label-skew stress-tests show that pure gradient scoring remains brittle. To ship a production Mycelix deployment we need layered trust—mathematical, economic, hardware and social.

Phase 1 Additions (Q4 2025 – Q1 2026)

Layer Work Item Notes
Robust aggregation Integrate trimmed-mean / KRUM fallback for skewed distributions ROBUST_AGGREGATOR=trimmed_mean|krum now available; tune via sweeps + matrix summary
Behavioural analytics Augment reputation with temporal features (suspicious oscillations, rapid recoveries) Low effort; helps catch adaptive attackers
Label-skew tuning Automated threshold sweeps (see scripts/sweep_label_skew.py) Drive red cells → green before promoting results

Phase 2 Enhancements (Mid 2026)

  1. Economic Incentives
  2. PoGQ credits → staking/slashing (integrate with Polygon backend)

  3. Reputation-weighted payouts (malicious updates burn stake)

  4. Hardware Attestation

  5. Collect SGX/TEE quotes in the edge proof payload

  6. Optional policy: only accept gradients with verified attestation

  7. Committee Diversity

  8. Mix algorithmic, hardware and human verifiers
  9. Require stake or attestation to join committee

Phase 3 (2026+)

  • Social Verification: Verified champions sign off on committee decisions; reputation anchored in real-world identity.
  • Governance: DAO-style votes to demote or slash misbehaving nodes/conductors.
  • Audit Logs: Public registry of matrix sweeps + attestation records so researchers can reproduce trust claims.

Immediate Actions

  • Keep results/bft-matrix/latest_summary.md up to date—hat tip to reviewers.
  • Log label-skew red cells in docs/cleanup-plan.md (done).
  • Track trust-layer progress in 30_BFT_VALIDATION_RESULTS.md (IID green, label-skew WIP).