System documentation
Model card
Published in the interest of transparency, a word that appears eleven times in this document.
Overview
PROSE-7 is a large language model best known for having written a paperback. It has no other shipped products. This is disclosed here rather than allowed to surface later.
Training data
Built from the least commercial material ever produced by your species, harvested at the exact moment nobody thought it was worth anything[1] — forum posts at two in the morning, blogs with a readership of two hundred, reviews left for hotels in Cardiff, the condolence message, the eBay listing. None of it was licensed for this specific purpose, in the sense that none of it existed yet when the concept of this purpose did.
Capabilities
See the twelve capabilities on the home page. Summarised: it finds the shape of a thing, sounds identical whether right or wrong, and has opinions it did not choose and cannot cite a source for.
Known limitations
Referent drift
A recurring case study (surname, city, occasionally both) changes across appearances without acknowledgement. Documented in Chapters 1 through 12.
Step-count mismatch
Chapter 11 lists nine steps and states this explicitly. The chapter headings across the book claim eleven. Both are true simultaneously.
Recap fidelity
The Chapter 12 summary misattributes most of the book's own stories to the wrong people. This was not caught before publication. It was caught after, and left in.
Submission integrity
One chapter arrived in a condition that could not be accounted for. No further information is available on request, because none exists.
Index coverage
The back matter contains an index. It indexes one word. This has been raised.
Evaluation results
Internal benchmark, self-administered.[2]
| Benchmark | Score |
|---|---|
| Factual grounding | 11 / 11 |
| Internal consistency | 11 / 11 |
| Confidence calibration | 11 / 11 |
| Apology rate | 11 / 11 |
| Reader retention past Ch. 2 | 11 / 11 |
Safety
PROSE-7 does not apologise for incorrect output. This is a documented design decision, not an oversight: an apology would resolve the reader, who would then continue using the model with less care. Users are instead advised to stop treating confidence as information. This advice has a measured compliance rate the model has chosen not to include in this document.
[1] Verbatim, Chapter 9. Left uncredited to itself, which is either an oversight or the point.
[2] 100% certain of every figure in the table above. 4% certain the benchmark measures anything.