Draft: reduce AI-cadence tells in thirteen-years book intro

[benbalter]

Draft for review — do not merge as-is. Conservative pass on the sections Fast-DetectGPT (Llama-3-8B) flagged as most machine-like. This is copyright-reserved book-intro material, so I only cut generic filler and antithesis tells and added specificity where I could without inventing biography.

Section scores (Llama-3, before → after)

section	before	after
What thirteen years taught me	66%	36%
Intentionality over proximity	62%	45%
A hub, not a headquarters	93%	85%

What changed

• Dropped "The opposite was true", "isn't just X—Y", "aren't just X. They're Y" antithesis tells.
• Cut "This flexibility stood in stark contrast to the traditional office culture of the time" (generic filler).
• Tightened the repeated "serendipity must be intentionally designed" point into "You don't lose the serendipity. You just have to put it on the calendar."

Note on "A hub, not a headquarters" (still 85%)

Left it essentially as-is. It's already concrete (Hack Houses, Destinations, cafés) yet scores high — a useful reminder that concreteness lowers the signal on average but isn't a guaranteed lever. Hacking a genuinely-concrete section to move a noisy number wasn't worth it.

Caveat

Whole-post % is a length artifact (criterion grows ~√N); section scores are the real signal. See #1954 for the full method writeup.

🤖 Generated with Claude Code