An option used to do exactly what it said. Now it may carry a check: one of
four attributes (Nerve, Wits, Charm, Resolve) against a set difficulty or
against a named person's own attribute, leaning on how they regard you, with a
success and a failure. The player sees only the odds in words ("a long shot",
"likely"), never a number. The roll is one draw from the run's generator, so a
run still replays exactly from its seed and choices.
A character may be a role with several authored candidates, one drawn per run,
or a group of two to four drawn into numbered slots. Everyone carries a name, a
description and the same four attributes as the player, so a replayed pack
meets a different office. Text names people as {partner} or {clerk}, and an
event about "a clerk" binds one present clerk for the turn. The draw and the
binding are saved with the run, and a save from before a pack drew a cast gets
the same one every time it is loaded.
All of it is generic and optional: Corporate Ladder and the Frontier use none
of it and play as before. The Counting House, a one-stage office about 1910
built from the story-to-pack catalogue, exercises it, and was tuned against the
conformance suite by measurement. DECISIONS #41. Version 0.4.0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018MhABSrqkwTdcybXmheDPr
The planning docs commit to one thing above all: adding content must never
require touching the engine. This lays the foundation that makes that true.
All game state is addressable by dotted path (money, skills.negotiation,
relationships.boss.likes_you, flags.took_loan), and conditions and effects are
generic operations over those paths. The schema in the planning docs had
`requires: { skill, min }` and `effects: { money, skills, relationships }`,
both of which hardcode the state shape into the engine; the first gate wanting
a relationship threshold or a flag would have meant an engine change.
There is deliberately no "random event" category. Every turn filters the whole
event pool by stage and conditions and draws by weight, so a routine turn, a
rare interruption and the hard-times branch differ only in their data.
Also here, neither in the planning docs but both cheap and load-bearing:
- A seeded, serializable RNG. Runs replay exactly from seed plus choices, which
makes tests deterministic and playtest reports reproducible. The generator's
position is one uint32 living in the save.
- A load-time content validator. Undeclared paths, unknown character ids, ids
that cannot be path segments, events that can never fire and stages with no
fallback event are all caught on load instead of forty turns into a game.
The turn loop is a pure function, so a saved game resumes into exactly the run
it left — covered by a test that plays thirty turns, plays the same thirty with
a JSON round trip in the middle, and deep-equals the two.
72 tests, no dependencies. Reasoning recorded in docs/DECISIONS.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VMSFHyVPitUoosW5wyEADj
