Data files

Four files in the repository are data the bot depends on. Two load into the database, one is compiled into the binary (and loaded), and one drives the evaluation. Pull requests that improve them are the easiest way to make the bot better.

data/aliases.yaml: nicknames

A flat mapping of nickname to canonical card name: bob: Dark Confidant, goyf: Tarmogoyf, t3feri: Teferi, Time Raveler. Aliases are lowercased on load and must match a current card name or face name case-insensitively. Nicknames that could mean several cards (the Tron lands, “karn”, “emrakul”, “sheoldred”) are left out on purpose. Ambiguity goes to the user as a “did you mean…?” prompt, never to a guess. Load with judge-ingest aliases data/aliases.yaml. With no file, aliases loads the copy compiled into the binary, which is what init and a container use, so an edit reaches an image only through a rebuild (or by mounting the file and naming it). The resolver also tries an alias with a trailing possessive stripped (“bob’s”) and as a suffix.

data/notes.yaml: nightmare cards

Hand-written Markdown notes keyed by card name, for cards whose interactions the rules text alone explains badly: Blood Moon, Urborg, Humility, and their kind. A note is injected into the model’s context whenever its card resolves. Notes are hints. The Comprehensive Rules still govern, and a note should cite rule numbers so the model can quote the rules rather than the note. Load with judge-ingest notes data/notes.yaml, or judge-ingest notes for the built-in copy.

data/categories.yaml: the taxonomy

The question categories the extractor classifies into, each with the CR sections that are always injected for it. crates/core/build.rs generates the Category enum from this file at build time, so editing it is a recompile and every match over the enum has to follow. The first category in the extractor’s output is required (an empty classification fails the model’s schema), and its sections lead the retrieved material.

eval/gold.yaml: the gold set

Adversarially verified rules questions with the cards they mention, the categories expected, the rule ids a correct answer must cite (expected_rule_ids), alternate ids that state the same fact (equivalent_rule_ids, keyed by an expected id), and a reference answer. Rule ids must be quoted ('614.12'). Unquoted, YAML reads 702.10 as a float and drops the zero, and the loader rejects it.

• judge-eval recall checks retrieval against it for free.
• judge-eval answer runs the full pipeline (about $2.50 for the set).
• rescore re-grades stored runs after an edit.

Extend it when adding capability, and re-verify rule ids on each CR release.

Reloading

aliases and notes are not part of the nightly refresh. Run their judge-ingest commands when the files change. The image carries binaries only, no data/. On a deploy host without a Rust toolchain, mount the file in: docker compose run --rm -v ./data:/data:ro refresh aliases /data/aliases.yaml.