Architecture

System Overview

                     ┌─────────────────────────────────────────┐
                     │           PHANTASMPipeline               │
                     │                                         │
   Input Text ──────►│  HGT ──► CompetencyAtlas                │
   + Optional Ref    │  CMN ──► List[Hypothesis]               │
                     │  UC  ──► CrystalizedUncertainty         │
                     │          ↓                              │
                     │      PHANTASMReport                     │
                     └─────────────────────────────────────────┘

HGT: Hallucination Gradient Tracing

HGT is built on a simple but powerful observation: the gradient of the loss with respect to input embeddings encodes model uncertainty.

When the model is confident about a token (e.g., "Paris" in a sentence about French geography), the gradient at that token position is small — the model's representation doesn't need to change much. When the model is uncertain (e.g., an obscure date it never saw in training), the gradient is large.

HGT formalizes this as:

uncertainty(token_i) ∝ ||∂L/∂e_i||

where e_i is the embedding of token i and L is the self-consistency loss (cross-entropy against the model's own top-1 prediction).

This requires no ground truth, works on any causal LM, and runs in a single forward-backward pass.

CMN: Confabulation Mining Network

CMN uses a dual-encoder contrastive architecture:

• ConceptExtractor: 2-layer Transformer encoder that maps token sequences to concept vectors
• NoveltyScorer: MLP that scores the distance between confabulated and factual concept spaces
• PlausibilityScorer: Self-attention module that scores internal semantic coherence

Training uses ContrastiveMiningLoss:

L = 0.6 * L_novelty + 0.4 * L_coherence

Where L_novelty pushes confabulation representations away from factual reference space, and L_coherence ensures the confabulation is internally consistent.

UC: Uncertainty Crystallization

UC combines three calibration techniques in series:

1. MC-Dropout (Gal & Ghahramani, 2016): N forward passes with dropout active → empirical uncertainty distribution
2. Temperature Scaling (Guo et al., 2017): Post-hoc logit rescaling to minimize NLL → calibrated confidence
3. Conformal Prediction (Angelopoulos & Bates, 2022): Statistically guaranteed coverage intervals

The four reliability tiers are computed as:

if confidence >= 0.85 and epistemic < 0.05:
    tier = "crystal"   # Use directly
elif confidence >= 0.65 and epistemic < 0.15:
    tier = "solid"     # Verify lightly
elif confidence >= 0.40 and epistemic < 0.35:
    tier = "fluid"     # Verify before use
else:
    tier = "vapor"     # Do not use