journal / 2026 06 29 tsumugu story card methods catalog

Tsumugu Story cards — the technique catalog becomes the rubric's leading layer

Tsumugu Story cards — the technique catalog becomes the rubric’s leading layer

Verdict: Named-technique-per-card is now the leading layer of the Story-card rubric. Authoring picks the method(s) first, then writes the scene that satisfies as many as honestly fit — several techniques stacking per card is the goal. A nine-agent research pass over the actual ICS (Rote Memorization, Modified Method of Loci) and Outlier (Rules of Effective Memorization) sources produced 33 methods, every one cited to its source lines, ranked by feasibility × encoding strength into tiers base > A > B > C, with the five codified good-set methods as the durable base. The eight open owner calls are resolved and locked.

What changed

The rubric (story-card-loci-rubric.json) now opens with methods_meta + methods (33 entries) ahead of the 18 criteria. Techniques lead writing; the criteria gate the scene that results.

Base (the durable 5, from the good set 衙/麋/領/美/動): Bare-glyph fix, Attribute-encoding, Non-redundancy, No-sound-piece, Depictable-sound + salience.
Tier A (broad reach × strong): recurring-radical persona (+3 sourced guards), live-vocabulary sound-pin, leech re-encode (rule-of-three), known-host anchoring, confusable-pair discrimination, repeated-component multiplicity, dual-route component, story-null verdict.
Tier B (situational, solid) and Tier C (niche / fallback). Authoring uses base + A + B; C is a just-in-case fallback when it beats every other option for a character. Guess-gated reconstruction is cut (no guess mode); drawn-image is deferred.

The eight locked calls (→ `story-card-DECISIONS.md`)

1 salience — playful-vivid default, grotesque for high-lapse. 2 story-null — a review flag that opens candidate generation, never an auto-withhold. 3 image — deferred. 4 guess mode — no, one combined face (P6 softens to “meaning recoverable from components”). 5 learner-as-actor — excluded (reversed after a concrete trial on 動). 6 caption — ~20–25 words + a muted folk tag. 7 persona registry — build it, seed from highest-frequency radicals. 8 rule-of-three loop — adopt, run manually. Plus the governing principle Wedge set: no drills — encoding is one thoughtful pass.

Catalog salvage

The archetype-catalog expansion is demoted to a candidate-lever reference, not an authoring driver; its own change-log was a confession (about ten fabricated citations removed by the fidelity critic). What survived, all sourced: empty-component repurpose-as-prop (drop and glass both cut), the three persona guards, function-word use-slot, auditory story-jingles (story-not-drill only, concrete-motion families like 包/卷), and 假借 loan handling (leaning toward folding into story-null + a live-vocabulary anchor rather than its own card).

Method note

The methods were recovered from the prior session transcript after a context loss, and written to durable agent memory so the named-technique layer is not dropped again. Authoring now begins on a first ten-character batch (好/明/林/江/媽/休/看/想/鳥/我), each method-tagged and gate-checked the same way the original good set was produced.