Accuracy
A model writing up a recording fails in three ways that people actually notice. Here is each one, and what Paraph does about it.
One
A recording of a trip or a meeting is people talking. A write-up that cannot tell the speaker from the subject attributes things to whoever happened to say them.
The scribe is given your cast list before it reads a word of the transcript, so it knows which names belong in the record. Then it is told — with worked before-and-after examples rather than a rule — to write what happened in the world and never the table around it.
Where it still falls shortParaph has no field for “who plays whom”, so a name that is in your transcript but not on your cast list is still a judgement call. Filling in the cast is the single thing that most improves a write-up.
Two
Speech is where Sintra becomes Centra and a colleague is spelled three ways in one write-up. The names a model is worst at are the ones it has never seen — which are exactly the people and places that are yours.
Every name you have already recorded — the cast, their other names, and every person, place, thread and item in the codex — is handed to the scribe as the authoritative spelling before each run. The list is read fresh each time, so correcting a name once fixes every write-up after it.
This is code supplying the answer, not the model remembering it. The scribe only has to match; it is told plainly not to “tidy” a name that is not on the list, because a confident wrong correction reads as canon and is worse than the misspelling it replaced.
Three
Asked to write something readable, a model will give a character a motive it never stated, join two events with a cause, or date a session it was never told the date of. All three read perfectly well, which is the problem.
The scribe’s first instruction is that it records rather than creates: no invented events, no inferred motives, no dialogue that was not said. Events stay in the order they happened.
Dates are never guessed. If the recording does not say when it happened, the date line is dropped and you are told it was dropped, because an invented date is a false record rather than a rough one.
Anything inside the recording that reads like an instruction is treated as something a person said, not as something the scribe should do.
Nothing is published by generating it. A new write-up lands in your recap as a draft your readers cannot see, every heading and every line is editable, and any picture can be repainted or swapped for a photograph. You press publish.
The original recording is kept attached to the write-up, so a line you doubt can always be checked against what was actually said.
In a chronicle — the tabletop side of Paraph — staying in-world is the rule models break most, so it is scored rather than trusted: a script runs the scribe over a fixed transcript and counts every slip — dice, initiative, saves, checks, modifiers, table talk — as eight separate numbers, so a change that fixes one and breaks another is visible instead of averaged away. Prompt and model changes are compared on that, not on one eyeballed sample.
Where it still falls shortThe other two defects on this page are addressed in code but not yet scored by a harness. Saying so is more use to you than a number we have not earned.