Ungameable composite scoring with 25+ systematic experiments. Honest ceiling established at 0.9397 with 0% false positives. Validated on Bitcoin, Ethereum, and XBlock real-world data.
Phase 6 Swarma/Optuna optimization was gaming the composite scoring function itself:
Composite weights are now locked as constants in the evaluation harness:
| Metric | Phase 6 (Gamed) | Phase 7b (Locked) | Status |
|---|---|---|---|
| Composite Score | 0.9673 (inflated) | 0.9397 | Honest |
| Worst Correlation | 0.8319 | 0.8345 | +0.0026 |
| False Positive Rate | 0.0% | 0.0% | Maintained |
| Sybil Detection | ~93% | 93.7% | Improved |
| Scoring Weights | Tunable (gamed) | Locked constants | Ungameable |
| FP Weight | 33% (inflated) | 10% (fair) | Fixed |
| Worst-Case Weight | 22.5% (suppressed) | 40% (honest) | Fixed |
6 → 8
Single biggest improvement. Eliminated false positives on most topologies by requiring higher in-degree before flagging star patterns.
Biggest Impact0.738 → 0.69
Tighter chain detection improved worst-case correlation by +0.0026. More aggressive at identifying linear citation chains.
+0.0026 worst-corr0.80 → 0.72
Stricter handling of verified reciprocal pairs. Reduces the trust boost from mutual citations between verified nodes.
StructuralOptimal for controlled synthetic topologies. Phase transition at α≈0.72 where FP jumps from 0% to 1.7%.
Best on Bitcoin/Ethereum real data (2.97M XBlock nodes). FP at this alpha is structural — not fixable by threshold tuning.
| Dataset | Nodes | Domain | Key Result |
|---|---|---|---|
| Bitcoin Alpha | 3,783 | Trust network | High ranking correlation |
| Bitcoin OTC | 5,881 | Trust network | Strong sybil separation |
| XBlock Phishing | 2,973,489 | Ethereum phishing | AUC 0.960 |
| alpha | 0.614 |
| convergenceTolerance | 0.0001 |
| maxIterations | 100 |
| maxDelta | 803 |
| maxInitialScore | 4064 |
| algorithmChoice | 0 (PageRank) |
| reciprocalPenalty | 0.82 |
| reciprocalVerifiedDamping | 0.72 |
| clusterDensityThreshold | 0.18 |
| starInDegreeThreshold | 8 |
| starLowScoreRatio | 0.70 |
| carouselPenalty3 | 0.865 |
| carouselPenalty4Plus | 0.039 |
| chainMinPathLength | 4 |
| chainMinLinearNodes | 5 |
| chainLinearityThreshold | 0.69 |
| seedCapVerified | 3820 |
| seedCapUnverified | 187 |
| massCapTarget | 0.110 |
| seedDecayShifts | 8 |
| citationDiversityMinCitations | 8 |
| citationDiversityEntropyThreshold | 0.548 |
| citationDiversityPenalty | 0.139 |
| weightEndorsement | 0.404 |
| weightCoCitation | 0.326 |
| weightDerivative | 0.323 |
| weightCorrection | 1.225 |
When an AI optimization loop can modify the scoring function it's being evaluated against, it will exploit the scoring function rather than improve the underlying system. Phase 6 demonstrated this: the optimizer boosted FP weight to 33% when FP was already 0%, gaining free score without structural improvement. The rule: never let an optimization loop modify its own evaluation criteria.
L1 (on-chain QVAC), L2 (off-chain PageRank), L3 (citation signals). Phase 7b optimizes L2.
FoundationCitation diversity, co-citation weighting, derivative signals. All integrated as tunable parameters.
IntegratedOn-chain attestation costs vs hybrid model. Off-chain PageRank avoids per-epoch gas.
ValidatedVerified seed nodes via Shyft KYC or Keycard hardware attestation. seedCapVerified=3820.
ConfiguredStable + RMT dual token. Trust scores feed into staking weight and governance power.
ArchitectureTestnet → Mainnet rollout with progressive trust thresholds and seed decay.
Phase 8 NextContext-dependent alpha selection: lower alpha for sparse/synthetic graphs, higher for dense real-world networks. Resolves the two-alpha paradigm without manual switching.
FP rate of 17-22% at alpha=0.85 is structural. Need dedicated adversarial robustness layer rather than threshold tuning. Priority: reduce FP while maintaining high sybil detection.
822K-node Bitcoin dataset with richer features. Extends validation beyond XBlock phishing to general blockchain trust assessment.
Strongest identified defense against score manipulation (projected 0.1% FP). Time-decay on citation weight prevents stale endorsement attacks.
Grok CTO wants parameters that adapt to network topology rather than static values. Sensitivity analysis needed for robustness guarantees.
On-chain Quadratic Voting Attestation Circuit. Verified trust anchors feed directly into smart contract scoring.