Catalog: mode-aligned staging surfaces (persuade/operate/read/experience), star ratings on approvals feeding challenger draw weights, family retirements, authoring strategy and territory guide, rework and breadth authoring rounds, composition mining from rejected worlds. Seed: six challengers (two per tier), --reroll chains, --mode staging filter, rating-weighted draws. New-work: Present/visualize/re-roll flow, image-gen requirement, register-neutral vocabulary. Pipeline: per-mode staging prompts with split frames, hero-from-board reference generation, render-safety guards. Labs: ratings UI, unrated filter, mode chips, composition approve-guard. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
5.2 KiB
World catalog authoring guide
How new concept-world rounds are authored, gated, and reviewed. Distilled from the full human sweep of 2026-07-20/21 (325 entries reviewed, 169 approved, all approved entries star-rated). The machine-readable core lives in skill/scripts/concept-ingredients.json under qualityBar (rejectIf, authoringStrategy); this guide carries the reasoning and the territory map.
The pipeline
- An authoring agent reads
qualityBar(includingauthoringStrategy) and the review file. Three-star approvals are positive exemplars; rejection and rating notes are negative space. - New entries merge as
pending. Nothing ships without human review. - The render gate runs before review: specimen board first, then the desktop hero generated with the board attached as binding reference (
scripts/generate-world-cards.mjs; the images/edits path keeps both images one system). - The reviewer decides in
/labs/worlds: approve or reject, star ratings on approvals (3 exceptional, 2 solid, 1 marginal), notes on anything instructive. Ratings feed challenger draws inconcept-seed.mjs(3-star doubles odds, 1-star sits out).
What wins
The winner-property test. Every candidate must be:
- Born-designed: the source is a produced 2D or display artifact with an existing graphic system, not a material, mood, or place. Even atmosphere-tier winners are secretly graphic (wax-print cloth, brick-build instructions, raku surface).
- Dense: the tradition immediately yields palette, materials, a type voice, several component roles, and a signature state change.
- Era-and-school specific: "1950s Blue Note session sleeve", never "record covers".
- New territory: the peak artifact of a culture the catalog has not touched. Proven seams saturate fast: the 2026-07-21 depth round scored 3/12 with 0 flagships because second-tier artifacts from mined veins read as near-duplicates ("too similar to others we already have"). Breadth-first beats depth-first.
- System-distinct: check the candidate against approved entries at the system level (palette plus type voice), not just by name. A different artifact with the same green-phosphor system is a duplicate.
Flagship share by tier after the full sweep: graphic 52%, interaction 55%, atmosphere 29%. Interaction display languages stay overweighted; atmosphere qualifies only through the intrinsic-pattern rule in rejectIf.
What loses
- Translation failures: the dominant rejection ("doesn't translate to interface"). Material worlds without an intrinsic 2D pattern system never recover, and re-authoring them fails again (rework hit rate ~33%, and only for taste fixes, never translation fixes).
- Too narrow (single prop, single color), too abstract, operations archetypes, generic categories.
- Render traps: brass or metal interface chrome reads cheap (skeuomorphism itself is fine; execution is the issue), non-Latin copy drift on non-Western worlds (interface copy stays English), AI-cliche motifs (Matrix glyph rain, recording dots), dated game chrome (game-born worlds are welcome at contemporary award standard), em dashes in rendered copy.
Cultural care
Skip living sacred, ceremonial, or community-owned traditions without an established commercial graphic lineage; skip highly religious material outright. Traditions enter cleanly when they already have a commercial or civic design history the way adinkra printing, thangka-informed diagram craft, hanafuda (a published card game), or azulejo (civic architecture) do. When in doubt, leave it out.
Territory map (unmined as of 2026-07-21)
One candidate per territory, always the territory's canonical peak:
- Print and publishing: Penguin/Pelican Marber grid, ligne claire comics, fotonovela, children's book schools (Scarry, Golden Books), Victorian trade cards, stamp design, marbled endpapers.
- Technical and scientific illustration: Haeckel plates, patent drawings, anatomical atlases, exploded-view manuals, Sanborn insurance maps, airline safety cards.
- Fashion as artifact (not draping): tartan clan registries, kimono/obi pattern grammar, sewing-pattern envelopes with tissue markings, Take Ivy catalog photography, sneaker-box and colorway-naming culture, techwear spec labels.
- Games and decks: tarot (commercial deck lineage), hanafuda, mahjong tiles, Tamiya model-kit box art with sprue diagrams, Game & Watch LCD language, casino chip and felt graphics.
- Global commercial traditions: Portuguese azulejo, heraldry, sonidero/cumbia posters, Jamaican soundsystem graphics, Tropicália, Ethiopian commercial iconography, Mayan codex facsimile publishing.
- Fine-art movements with systems: Vorticism, Precisionism, Le Corbusier polychromy, Barragan color walls.
Mark territories off as rounds mine them; a mined territory moves to the saturation gate.
Round mechanics
- Rounds stay small (~12) and are treated as experiments; expected hit rate is 25-40% now that the initial canonical harvest is done. Read the verdicts before the next round; every rejection note is calibration.
- Standing reviewer requests: more non-Western systems, more cassette/VHS-era systems.
- Retired families never receive new entries: machines-contraptions, food-potioncraft, memory-ruins, festivals-public-life.