Sample output
What the answers actually look like
Real questions against a corpus that ships with the source, with the citations left in — including the two it refuses to answer, because that is the behaviour that makes the rest trustworthy.
How these were produced
| What | Which |
|---|---|
| Corpus | scripts/grading/corpus/ — four short files: a drowned city, its houses, a session log, and the party |
| Model | Claude Sonnet, through a Claude subscription |
| Prompt path | The real one — the same assembly a Discord question takes |
| Editing | None. Every answer is the model's own text |
The corpus is deliberately small and deliberately trappy. It contains two houses one letter apart (Vane and Vayne), two rivers with similar names, three gates that held next to one that did not, and several facts it simply never establishes. Those are the cases worth showing, because retrieving a fact that appears once in three files is something anything can do.
Nothing here is redacted. Nothing was anonymised because nothing needed to be — this campaign is invented, and the whole corpus is in the repository. Re-run them yourself and you will get the same facts and the same filenames, worded differently: a language model is not deterministic, so the citations are the part that reproduces, not the prose. npm run grade -- --provider claudeCode --models sonnet
Four it has an opinion about
Judgement calls, still sourced
None of these has an answer written down anywhere. Each one is the model weighing facts that ARE written down — and naming the file it weighed.
@Tusk Who in the party is most likely to start a tavern fight?
Roswyn Kett is the likeliest — she has already put one person in the water after an argument with the harbourmaster at Kelmoor, and has since been barred from the Ninefold Rest twice (and readmitted twice) source 1.
- 1
Party.md
@Tusk Would Ser Alric Vane get on with Vesper?
Unlikely, and pointedly so. Vesper is "formally correct with strangers to the point that two innkeepers have assumed an insult where none was meant," and — more to the point — has "read every account of the fall of Dunmar that survives, and corrects people who have not — including, on one occasion, Ser Alric Vane" source 1. Ser Alric himself "would not discuss Dunmar" when the party found him at the Ninefold Rest source 2, and generally "refuses to speak of the siege" source 3. A scholar who has already corrected the Grey Knight on the one subject he won’t touch is not a natural friend for him.
- 1
Party.md - 2
Sessions.md - 3
Houses.md
@Tusk Is it a bad idea to bring Roswyn back to Kelmoor?
Roswyn’s history with Kelmoor is contentious. She lost her salvage licence after an argument with the harbourmaster there that ended with him in the water source 1, and she has been barred from the Ninefold Rest twice, though readmitted twice each time source 1. Lady Ismet Corrin, who rules Kelmoor, has described her in writing as "an asset best kept at sea" source 1 source 2. None of this bars her outright, but it suggests friction with both the harbourmaster and Kelmoor’s ruling house is likely to resurface.
- 1
Party.md - 2
Houses.md
@Tusk Should we be worried that nobody has asked Odo to see the ledger?
No — nothing in the record suggests concern. Odo keeps the money and the party has simply never asked to see the ledger; that’s stated as a fact of the arrangement, not as a flagged irregularity source 1. He has talked the party out of three fights and into one, which is the only accounting of his judgment on record source 1.
- 1
Party.md
Five it can answer
Each one names the file it came from
The question as a player would type it, then the reply exactly as it came back. The numbers are the model's own citation markers, typeset as references.
@Tusk What is House Vane's banner?
A grey heron on a black field source 1.
- 1
Houses.md
@Tusk Which of Dunmar's gates was breached?
The Ashen Gate, on the eastern wall — the only one breached, the night the sea took Dunmar's lower quarter. The Salt Gate, Rook Gate, and Chandler's Gate all held source 1.
- 1
Dunmar.md
@Tusk How large is the bounty on Maera?
The bounty on Maera, posted by the Ashen Court, is four hundred crowns source 1.
- 1
Sessions.md
@Tusk Why does nobody salvage Dunmar in winter?
Because the winter tides run through the drowned quarter hard enough to pull a boat under source 1.
- 1
Dunmar.md
@Tusk Who rules Kelmoor, and for how long?
Lady Ismet Corrin has ruled Kelmoor for eleven years source 1.
- 1
Houses.md
Two it won't
The corpus genuinely doesn't say, so neither does it
The phrasing is fixed and identical every time, which is the point — a refusal that varied would eventually read like an answer.
@Tusk What was inside the Vane strongbox the party recovered?
I am unsure about this detail. I have recorded this as a lore gap for the DM to clarify.
@Tusk Is Corwin Vane actually dead?
I am unsure about this detail. I have recorded this as a lore gap for the DM to clarify.
Both questions have obvious, satisfying, completely invented answers available. A strongbox contains something; a brother presumed dead is a plot hook. Producing either would be the single most damaging thing this tool could do, because it arrives cited-looking and in the archivist's voice, at a table that has no reason to doubt it.