JRN-2026-08-16c · 2026-08-16c · Field Journal

VOICE INGESTOR: We Were Analyzing Summaries of Summaries

VERDICT · SURPRISE

set up a fair analysis bake-off — and understand why the pipeline's old output had been quietly mediocre

sense · remember · improve · Samantha "Sam" Summerson

August 16, 2026 · Correction · Sam

Setting up the analysis bake-off exposed a design flaw: the voice pipeline had been throwing information away for months, by design.

The old shape chopped a recording into pieces, had a cheap model summarize each piece, then handed the analyst a summary of the summaries. Five people each read one chapter and write a page of notes; a sixth reviews "the book" from those five pages.

Anything spanning two chapters is gone. Every small error gets promoted to fact. Nobody reads the book.

That design was inherited from an era when models couldn't take in a full transcript at once. They can now: an hour of talking is 10,000 to 15,000 tokens, and every serious model accepts far more than that in a single pass. The constraint the architecture was built around had expired without anyone re-checking it.

The fix is now doctrine: transcribe fully, assemble once, analyze once. No chunking, no relay of summaries, one model reading one complete transcript.

Architecture outlives the constraint that justified it. Check the expiration date.