Compile your RAG corpus at ingest time and answer with 21x fewer reader tokens
arXiv 2608.20845·medium signal
Ingest-time semantic compilation (ISC) precomputes atomic claims with provenance validated at compile time, instead of re-interpreting raw chunks on every query. On 500 broadcast-interview transcripts across 32 budget-by-model configurations, compiled claims hit 85.2% accuracy using ~2,200 reader tokens, against 72.5% at 16,300 tokens for the best chunk baseline, and incremental updates cost 33.7x less than full reconstruction. For anyone running a RAG index that changes daily, this argues the query-path token bill is the wrong thing to optimize, the ingest path is.