LSAP-1 — an instrument for measuring prose
A pilot protocol that scores a passage of fiction on thirty anchored axes — and a generation engine firewalled from the analyzer, so measurement can never quietly become imitation.
Can you measure prose the way you measure a signal? LSAP-1's answer is partially — and here is exactly how partially. It turns a segment of fiction into thirty anchored numbers across a handful of structural dimensions, rated by LLMs under a fixed protocol, over a purpose-written original corpus.
What's technically interesting
The analysis↔generation firewall: the instrument that reads and the engine that writes are architecturally forbidden from feeding each other — an invariant enforced in code and covered by its own tests, not just stated in the docs. The point is that a measurement system for literature must never quietly become an imitation system. Around it: structured-output raters, JSONL + git-diffable storage, and a reliability pipeline (agreement, correlation, PCA) over a two-rater, 30-segment pilot corpus.
Why it holds up
It publishes its own unflattering numbers: redundant axis pairs the correlation analysis exposed, a first principal component that eats much of the variance, and the plain caveats — a pilot-sized corpus, LLM raters with no human baseline yet, the generative compiler still unbuilt. The instrument is a hypothesis, and it's scored like one.