> IntegriRef: Five-Layer Reference Integrity Verification
A five-layer (L0-L4) citation verification engine: L2 semantic claim verification achieves 93.5% accuracy (exceeding human agreement at 89.1%), while L3 graph anomaly detection and L4 Bayesian risk scoring have no commercial equivalents. 100% hallucination recall at 0% FP rate.
// TABLE_OF_CONTENTS
IntegriRef: Five-Layer Citation Verification¶
The Stack¶
| Layer | Function | Performance | Competitors |
|---|---|---|---|
| L0 | Existence + hallucination | 336K refs/s | scite.ai |
| L1 | Citation intent (Support/Contrast) | 86.0% F1 | scite.ai (29% F1) |
| L2 | Semantic claim verification (NLI) | 93.5% (DeBERTa-v3) | None |
| L3 | Citation graph anomaly (7 types) | CIDRE + Benford | None |
| L4 | Bayesian risk scoring (18 signals) | Calibrated posterior | None |
Golden Benchmark (58 cases)¶
- Hallucination: 100% recall (14/14), 0% false positive (0/18)
- Retraction: 73.9% recall (17/23) vs CheckIfExist 26.1%
- Chimera: 100% (3/3) vs CheckIfExist 33.3%
- Tradeoff: 38.5s vs CheckIfExist 2.3s (17x latency for dramatically higher detection)
Why Naive Bayes Wins for Signal Fusion¶
| Method | Hallucination Recall | FP Rate |
|---|---|---|
| Naive Bayes + Kill-shot | 100% | 0% |
| Grouped Bayesian | 50% | 0% |
| Dempster-Shafer | 100% | 100% |
The kill-shot override is decisive: when phantom_doi fires (syntactically valid DOI that doesn't exist), score is forced to CRITICAL regardless of other signals.
Cascade Early Stopping¶
67.2% of references can skip L1-L3 when L0+L4 score is decisive, with zero accuracy loss. All skipped cases shifted upward (conservative direction).