Overview / Colophon What it took to create this project
One last bit of trivia, for anyone curious what producing a project like this takes.
Everything here — the data collection, the experiments above, and the site itself — was built in
conversation with Claude Code (Fable 5), an AI
coding agent: a human typing instructions and reviewing the results, the agent writing the code,
running the collections and keeping the books. The numbers below, measured from the agent's
session logs by ccstats
62 h of active work, over18 active days
574 prompts typed by the human
68,071 words typed 395,530 characters typed
23.0 M tokens of model input + output1.59 B including cache
389 K tokens of input22.7 M tokens of output
613 subagent launches15,821 tool invocations
3,175 web pages fetched 2,178 web searches
$3,023 what the tokens would cost at API rates
The cost figure is hypothetical: it prices the tokens at the vendors'
pay-as-you-go API rates. The project actually ran on a fixed-price Claude Max 20x subscription, so
that is what the work would have cost via the API, not what was paid. The headline token count
excludes prompt-cache reads — context re-served to the model, which dwarfs everything else without
representing new work; the "including cache" figure counts them. Subagent launches are the agent delegating subtasks to fresh copies of
itself — all of the Section 14 research ran this way: a fresh copy knows only what its own
instructions tell it, which is what made running those agents blind possible.