Evidence Attribution

The foundational step of decomposing source data into discrete, addressable units that can be individually cited. This moves beyond treating documents as monolithic blocks.

• Token-Level: For text, attribution points to specific words or subwords.
• Superpixel/Region: For images, highlights the exact pixel clusters influencing classification.
• Time-Slice: For audio or sensor data, isolates the precise temporal window of relevance.
• Node/Edge: For graph data, identifies the critical entities and relationships driving the prediction.

Establishes a non-negotiable, one-to-one mapping between a claim in the generated rationale and its origin in the input. This eliminates 'plausible but fabricated' justifications.

• Pointer Networks: Architectures trained to output specific input indices rather than generating free-form text.
• Hard Attention Masks: Forces the decoder to only attend to verified source segments during explanation generation.
• Contrastive Attribution: Highlights the evidence that differentiated the chosen prediction from the runner-up alternative.

Quantitative methods to ensure the highlighted evidence genuinely caused the prediction, not just correlated with it. This distinguishes faithful rationales from plausible rationales.

• Input Erasure Tests: Systematically removing the attributed evidence and measuring the resulting drop in prediction confidence.
• Sufficiency Metrics: Verifying that the cited evidence alone is enough to reproduce the original model output.
• Comprehensiveness: Confirming that all necessary information for the decision is captured within the attribution.

Synchronizing evidence across heterogeneous data types to provide a unified justification. A diagnosis might cite both a specific region in an X-ray and a symptom in clinical notes.

• Cross-Modal Attention: Aligning visual regions with textual phrases to create joint evidence maps.
• Temporal Alignment: Linking a video frame with the corresponding timestamp in an audio transcript.
• Unified Provenance Graphs: Constructing a single graph that connects evidence nodes from text, images, and structured databases.

The presentation layer that translates machine-readable attribution indices into an intuitive format for human review. Raw token IDs are useless to a compliance officer.

• Saliency Overlays: Heatmaps rendered directly on the original document or image.
• Inline Citation Tooltips: Highlighted text spans that reveal the source passage on hover or click.
• Confidence-Anchored Evidence: Visually encoding the statistical weight of each evidence piece alongside its location.

A critical evaluation dimension ensuring that the content of a generated rationale does not contradict established real-world knowledge or the specific source data provided. Evidence attribution fails catastrophically if the cited evidence is misrepresented or contradicted by the explanation text.

• Intrinsic Check: Verifies explanation aligns with the cited source segment
• Extrinsic Check: Validates claims against external knowledge bases
• Contradiction Detection: Flags mutually exclusive statements within a single rationale

Factual consistency transforms evidence attribution from a pointing mechanism into a truthfulness guarantee.

The degree to which a generated rationale accurately mirrors the true computational logic used by the model to arrive at a prediction. This is the ultimate measure of evidence attribution quality—a faithful explanation is one where the cited evidence genuinely drove the decision, not just correlated with it.

• Simulatability: Can a human use the explanation to predict the model's output?
• Sufficiency: Does the cited evidence alone produce the same prediction?
• Completeness: Does the explanation capture all influential factors?

Faithfulness is the gold standard that separates genuine evidence attribution from post-hoc storytelling.