Skip to case
TWILIGHT TECH LABSTHE ENGINEERING PROJECT BOOK

ENGINEERING INDEX / ENTRY 19

CASE 19 / MEDIA INGESTION / SOURCE STRUCTURE

Keep the page attached to the passage.

A listening edition needs manageable spoken passages, but blind splitting can sever headings, page references, and dialogue.

Source project: Twilight ReaderFocused extraction tests passedReviewed September 25, 2026
01 / THE ENGINEERING QUESTION

How can a document become narration chunks without losing where each passage came from?

02 / THE SYSTEM

Where the handoffs happen.

The extractor tests preserve page ranges and a heading, can flush chunks on page changes, keep dialogue and abbreviations together, reject table-like false headings, and read PDF metadata.

  1. 01Extract blocks
  2. 02Detect headings and sentences
  3. 03Chunk by source page
  4. 04Retain metadata
03 / THE CLAIM LEDGER

What the record earns.

Six PDF extraction/chunking checks passed.

Observed receipt
On September 25, 2026, six Reader extractor tests passed within a seven-test extraction/router batch.
Source and scope
Twilight Reader extractor and tests/test_reader_extractor.py; Reader source/tests untracked in dirty checkout with base HEAD efd4149.
04 / THE LIMIT

What this does not prove.

Fixture PDFs and hand-built blocks do not prove accuracy for scanned documents, every layout, or copyrighted source material.

05 / NEXT PROOF

Where the work goes next.

Build a rights-cleared mixed-layout fixture set and score extraction against human-reviewed source passages.