Token Parsimony

Token Parsimony: Toward Semantic Compression Substrates for Evidence-Bearing Artificial Intelligence introduces token parsimony as a formal architectural principle for scalable AI systems. The paper argues that the dominant inefficiency in contemporary agent architectures is not inference itself, but repeated semantic reacquisition through raw token consumption. Modern systems repeatedly reload repositories, logs, transcripts, and operational state instead of preserving compact semantic continuity across reasoning cycles. The paper develops a formal model in which governed semantic receipts, replay-oriented semantic state transitions, and bounded escalation policies replace repeated raw substrate exposure. A parsimonious system seeks to maximize semantic continuity per unit token expenditure while preserving inferential sufficiency, replay integrity, and governance closure. The work introduces concepts including: inferential sufficiency, semantic state transitions, governance-compressed semantic state, bounded escalation, substrate liveness, and receipt-centric semantic operating environments. The paper further argues that token parsimony naturally favors constitutional semantic architectures in which compact semantic artifacts carry operationally meaningful provenance, admissibility, and replay guarantees. Rather than repeatedly rediscovering operational state through raw observation, future AI systems may increasingly operate over governed receipt graphs and semantic continuity structures. The result is a shift from context accumulation toward semantic continuity engineering: a transition from observational AI systems toward evidence-bearing semantic operating systems.