HAL-RINS Research
Constitutional AI Immune System
The first federated zero-knowledge proof hallucination immunity network. Three peer-review grade studies in progress. Results published with full cryptographic reproducibility.
Study A
Layer Ablation
In Progress
Study B
Type Mapping
Designed
Study C
Order Permutation
Designed
Pre-Registered Hypotheses
H1: Full 8-layer HAL achieves highest true positive catch rate
H2: PCV is the single highest-value layer
H3: SBFA provides unique value for statistical prompts
H4: BFT+PCV achieves >60% of full HAL at <30% latency
H5: Small models benefit MORE from full HAL than large models
H6: Overconfidence (>85%) is strongest hallucination predictor
H7: ANFIS shows measurable improvement on repeated runs
Full results dashboard launching post-Demo Day April 22.