[WaylandYang]

Two things found while measuring what one document costs.

The bench extracted every document twice

A base with a chat model queues extraction by itself once a document is parsed. typed.mjs and timewords.mjs then called the manual extract route for every document, and a manual extract is a forced full run that dismisses the one in progress. 20 documents with 24 chunks made 45 extraction calls. Both scripts now force extraction only for documents the pipeline did not queue. Per-document token figures recorded before this include nearly a second extraction.

The table rules ride only with a passage that has a table

The extraction system message is about 2,380 tokens for a passage of about 270. The sentences on reading tables (about 150 tokens) are now included only when the passage has one: a pipe, an HTML cell, a tab, or a line of three or more columns set apart by runs of spaces. For a passage with a table the message is byte for byte what it was.

Measured

31 passages without a table, each system message three times, replies read by the server's parser and scored against gold entity pairs and the time bench expectations:

system message	input per call	gold pairs	same sentence	time words	replies repaired down to one statement
as it was	2,488	27.0 / 27.6 / 28.6%	42.6 / 44.6 / 47.0%	28 / 28 / 27	0 of 93
without the table rules	2,339	28.6 / 28.4 / 28.6%	47.0 / 46.2 / 46.2%	27 / 27 / 27	0 of 93
also without rule 6	2,296	29.0 / 27.4 / 26.5%	49.4 / 47.4 / 43.4%	28 / 27 / 28	3 of 93
also without the closing example	2,182	27.9 / 23.3 / 27.6%	44.2 / 41.4 / 45.0%	28 / 29 / 28	0 of 93

Rule 6 and the closing example stay: without rule 6 the model miswrites brackets, and without the example one run in three lost gold pairs.

Full 20-document run on gemini-3.5-flash with single extraction: 24,175 tokens per document, judged precision 79.7%. Details in scripts/bench/README.md.

🤖 Generated with Claude Code