Score a whole novel by emotion, one passage at a time

01 · Role in the system

What Jev does here

Each passage becomes a fitted state carrying ten parallel typed questions: a 0-3 score for each of nine named emotions, plus overall intensity. Jev never sees pixels and never writes prose; it returns scores with calibrated probabilities, and the app draws one feathered row per passage, each emotion a curtain whose width and brightness follow its score. The author reports the whole 601-passage read - 6,010 judgments - finishing in 24.7 seconds for about three cents, with six requests in flight and a duplicate hedge fired when one is slow. When the book is done the strip is the book's emotional weather, exportable as a poster.

Six thousand judgments, one shape, repeated

Book Aurora does one thing six thousand and ten times: every roughly 90-word passage of Frankenstein becomes ten parallel typed questions — a zero-to-three score for each of nine named emotions plus overall intensity. The uniform shape is what makes the economics work: 601 passages judged in 24.7 seconds for about three cents, six requests in flight, and a duplicate hedge fired when a request runs slow.

The model never sees pixels and never writes prose; it returns scores with calibrated probabilities, and deterministic code owns rendering, concurrency, hedging, and timeouts. The result — the book's emotional weather as an exportable aurora strip — is a visualization of judgments a reader can spot-check passage by passage.

Author-reported numbers, honestly framed

The 24.7-second and three-cent figures come from the author's own run and were not independently reproduced at verification time — the record says so in its metric conditions and limitations rather than letting the impressive number float free.

The engineering boundaries are also stated: requests can route through OpenRouter or the native API with the README pricing the gateway path, the server binds to localhost and refuses cross-site reads because every read costs money, and only public-domain texts ship with the repository. As a demonstration of dense parallel scoring, the record doubles as a template for any passage-level document analysis.

02 · Control boundary

Where Jev sits

Passages go out as fitted state with ten parallel score questions; deterministic code renders the returned scores as the aurora and owns concurrency, request hedging, and timeouts.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• The 24.7-second and three-cent figures are author-reported from the author's own run and were not independently reproduced here.
• Requests can go through OpenRouter (model typesafe/jev-1.13) or the native TypeSafe API; the README prices the OpenRouter path.
• Only public-domain texts ship with the repository; the server binds to localhost and refuses cross-site reads because every read costs money.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• dani1005 ↗Community · github · public · checked 2026-09-20