How session scores are calculated
Each prompt is graded on what its kind of task needed, and your ROIP is a weighted average of those grades.
Your ROIP is built in three steps.
The kind of task
Jev sorts the session into one of 8 categories and 43 task types, such as Email & Message, Debug or Research. The type decides which components are Core, which are Situational and which aren't used, and how much each weighs.
Each turn's grade
Each prompt's role decides what it's graded on (see Turns and the session score). Its turn grade, from 0 to 100, is the weighted share of the components it needed that it had. A component counts as had when this prompt, a file attached to it, or an earlier turn of the same task gave it. It earns full credit, half when the grader isn't sure, or none.
A prompt is graded on what it and the conversation before it gave the AI. The grader doesn't look at how the AI answered that prompt. Whether a reply landed or was corrected is shown beside each turn but doesn't change the score, and a Correction is graded like any other turn, even when the AI made the mistake.
The session
The session's ROIP is a weighted average of its turn grades. Depending on the task type, the first New Ask counts 3×, every New Ask counts 2×, or every turn counts the same. In Role-Play & Games and Rehearsal, in-character Extensions count 0. A Close isn't graded.
If the session switches to a different kind of task, it's split into parts. Each part is graded by its own type, and the parts are combined by how many graded turns each has.
A worked example
An Email & Message session, where the first New Ask counts 3×: a first prompt that gives everything an email needs grades 100. A follow-up that only says “that's wrong” gives no direction, so that Correction grades 0. The session's ROIP is (3 × 100 + 0) ÷ 4 = 75. A Correction that said what was wrong and what to change instead could have graded 100, and so could the session.
Not scored
Casual Chat and Emotional Support, the same long template reused with one word swapped, attempts to get around the AI's rules, unsafe requests, sessions mainly in a language other than English, prompts written to steer the grader, and sessions whose kind of task Jev can't tell with confidence.