The engine phase: rebuilds measured, stamped, and made incremental

20 July 2026 Β· Nabu news

The library tripled in a week, and a full rebuild grew to hours. Phase 36 answers with measurement before optimization: every rebuild now times itself per source and per stage, and the first instrumented run settled the question β€” the cost is dominated by database insertion, not parsing.

The deeper change is the derivation stamp. Each source’s derived rows are a pure function of four inputs β€” its canonical bytes, its parser code, the fold rules, and the schema β€” and every load now records a fingerprint of all four. nabu rebuild --incremental skips any source whose fingerprint is unchanged and re-derives only the dirty ones, with the full rebuild remaining the reference implementation, pinned equivalent by test. A frozen source costs nothing to keep; a parser fix re-derives exactly the shelves it touches.

The recalibration ruled in Phase 35 also executed: the quotation- detection thresholds in parallels are now corpus-relative β€” the old absolute cutoffs, tuned when the library held 3.8 million passages, had drifted to a sixth of their intended strength against 24.4 million. The reference pages refreshed to the settled census, recording a changing of the guard: Literary Chinese is now the library’s largest language, at 13.0 million passages. Numbers as of 2026-07-20.

← All news  Β·  Atom feed