Provenance-Conditioned Attention: Source-Type Gating for Epistemic Reasoning in Transformers

Standard transformer attention treats all tokens as epistemically interchangeable, computing relevance purely through semantic similarity of query and key vectors. This architectural design discards a critical signal in source-critical domains: the provenance type of the information being attended to. We introduce Provenance-Conditioned Attention (PCA), a mechanism that augments scaled dot-product attention with a learned, low-dimensional source-type gating channel. Each token carries a provenance embedding indicating its epistemic category (e.g., testimony, archival record, scholarly analysis), and a compatibility function over these embeddings multiplicatively gates attention flow. We present three architectural variants: multiplicative gating, additive score fusion, and head-partitioned attention. PCA adds approximately 0.01% additional parameters at target model scale and recovers standard attention as a special case. We validate PCA on three synthetic benchmarks of increasing difficulty across six training scales (500–20,000 examples) with five random seeds per configuration. On single-source retrieval, PCA matches a segment-embedding baseline from 2,000 training examples onward. On a compositional multi-hop task requiring simultaneous attention to two source types, all PCA variants achieve perfect accuracy (1.000 ± 0.000), compared to 0.921 ± 0.158 for segment embeddings and 0.181 ± 0.010 for standard attention. PCA addresses a gap orthogonal to recent work on epistemic uncertainty in attention: not "how certain am I?" but "what kind of knowing is this?"