A game master who uses language models to prepare sessions of a long-running tabletop
role-playing campaign needs grounding documents: a current world state, campaign state, party
record, planning notes and per-character dossiers, distilled from dozens of recorded sessions and
fed to the models as context. Producing these documents is easy to start and hard to get right,
because the details that matter are exactly the ones generative models get wrong: who did what,
in what order, which name denotes which entity, and whether something happened at all. Over
seven months (March–October 2026) and six campaigns of up to 70 chapters, one practitioner
built, measured and retired twelve distinct approaches to the problem. We reconstruct them
from the git history, design documents and experiment write-ups of the repositories involved,
and organise them by two axes: the substrate a pipeline reads (narrated chapter prose, atomic
extracted facts, retrieval indexes, or the structured session summaries the GM already reviews)
and who makes the precision decisions (ordering, attribution, identity, scope). Approaches
that let a model make those decisions failed in ways that were silent and that compounded:
a consolidated extraction shrank a planning document from 513 to 109 lines; an atomic-fact
ensemble reported a slain antagonist as alive, and 98% of its 15,465 facts were seen by only
one sample, so agreement could not serve as a filter; a knowledge-graph library stored 1.18
characters of generated text per character of source and answered “what was the most recent
session” with session 8 of 11. The approach that the GM finally promoted parses the scene order,
entity sections and memorable moments already authored in reviewed session summaries, with
no model call (67 summaries, 409 scenes, 868 dossiers, byte-identical across runs), and uses a
model only to render the final prose; it also flagged 51 claims in the GM’s own tracking files
as unsupported by the record. We distil the history into ten cross-cutting findings and a set of
design rules for grounding long-horizon generative work.
The full paper can be found here – https://drive.google.com/file/d/1USvYt7moAUf8GbDzUYwW33I9HOsTu39D/view?usp=sharing
Leave a Reply