An evaluation-first foundation for an assistant that converts The Wandering Inn's diegetic Class/Skill system to Pathfinder 2e. It turns the existing written rules into a small, testable software contract before any model is trained or connected.
src/grand_design/domain.py: typed input/output contracts for annotated conversion examples.src/grand_design/rules.py: the deterministic rules that must remain true regardless of the eventual model.src/grand_design/model.py: aConversionModelprotocol and a transparent rules-based baseline.data/: version-controlled JSONL records for source facts and evaluation cases.tests/: regression tests for established rulings and the baseline.
The model deliberately does not scrape websites or reproduce source prose. Add only concise, paraphrased facts from the project-approved wikis, with a source URL and a reviewer. See data/README.md.
Requires Python 3.10 or later and no runtime dependencies.
cd C:\Users\parez\code\GrandDesAI
python -m unittest discover -s tests -v
python -m grand_design.cli validate-data
python -m grand_design.cli evaluate
python -m grand_design.cli corpus-report
python -m grand_design.cli convert --input examples\innkeeper.jsonOptional editable install exposes the grand-design command:
python -m pip install -e .
grand-design evaluate- Add a reviewed, paraphrased source fact to
data/classes.jsonlordata/skills.jsonl. - Add a human-approved expected output to
data/evaluations/known_conversions.jsonl. - Run
validate-dataandevaluate; the deterministic baseline establishes a reproducible floor. - Implement an adapter satisfying
ConversionModelfor an LLM, classifier, or fine-tuned model. - Keep the rules engine as a guardrail: flag unresolved cases instead of inventing permanent rulings.
The current fixture set is intentionally small. It verifies the known Zel Shivertail power-tier ruling and the one-primary-class/one-archetype limit demonstrated by Klbkch. Expand it with held-out examples before judging a learned model.
data/ability_scenarios.jsonl contains original, deliberately unusual situations split into training and held-out evaluation examples. Each ability names a Pathfinder comparison anchor and declares its character level, spell-rank ceiling, action cost, target shape, duration, save posture, narrative trigger, and non-negotiable guardrails.
data/pf2e_power_bands.json is the hard limit: it advances the spell-rank ceiling every two character levels, and a record cannot use a spell-rank or action budget above its exact envelope. The validator also rejects Tier 3 abilities that lack a narrative trigger. Read data/PF2E_SOURCES.md before adding records; verify the anchor against current Archives of Nethys rules instead of copying rules text into the dataset.
Use evaluate-abilities --predictions path\to\predictions.jsonl to score a model against the held-out ability records. Each prediction must provide its evaluation id, proposed expected_tier, spell_rank, actions, and narrative_trigger. The scorer reports exact agreement and tier accuracy, then marks a prediction safe only if it passes the same PF2e envelope validation as the corpus.
data/class_scenarios.jsonl adds 24 original character concepts (16 train / 8 held-out eval) covering martial, magic, social, crafting, mobility, investigation, leadership, prestige, and world-category concepts. Every record identifies feat-chain evidence, the PF2e chassis, the exact level-band calculation, and whether a secondary Class is actually regular enough to justify its one archetype. The validator rejects unsupported chassis, incorrect power-tier scaling, unsupported secondary archetypes, and missing guardrails.
foundry-module is a Foundry VTT 12-13 module for PF2e worlds. It validates approved conversion JSON, stores it on Actor flags, and exposes an API for later automation. GAME_DESIGN.md defines the campaign loop, while world-map.md documents the original zoomable atlas scaffold and how to replace it with a licensed map asset.
Approved Foundry entries include tags and lineage metadata. The module automatically adds approved Grand Design Classes and Skills as custom PF2e feature Items without overwriting the Actor's normal PF2e chassis; later combinations and upgrades retain all source IDs, inherited tags, and the recorded rationale.
convert accepts a reviewed JSON character profile and prints the project's per-character documentation template. Skills require two explicit human judgments, matching the established checklist:
changes_identity: the ability fundamentally changes what the character is.maps_to_existing_mechanic: an existing PF2e feat or spell has the same mechanical effect.
This forces uncertain source interpretation into the input review step. The baseline then assigns a tier, preserves the one-primary-class/one-archetype limit, recommends a reviewed PF2e chassis when one exists, and flags choices that still need a person.
For a multi-Class character, mark exactly one Class as is_primary: true and, only when it is regularly relevant at the table, one different Class as is_secondary: true. The prototype never promotes an arbitrary second entry into an archetype.
The human-readable source of truth remains:
wandering-inn-pf2e-conversion-rules.mdAI-continuation-prompt.mdwandering-inn-dm-starter-kit.docxwandering-inn-player-guide.docxplayer-concept-gallery.mdfor spoiler-light, original player-concept examples
When a machine-readable record conflicts with those documents, correct the record and add a regression test.