Tsumugu: authoring opens, encoding review lands (2026-06-11, later)

Same day, third act. The morning gave the corpus custody; the evening opened production: 40 new entries authored and approved (10 backbone words, 30 character entries — every character behind both word batches), the corpus at 53, and a structured review of the three source documents (encoding deep-dive, OUTLIER-CONDENSED, ICS-CONDENSED) that turned into two amendment rounds on the entry contract.

Decisions

Production model proven at the 10-unit gate. Wedge’s ramp — 10 for review, 10 more, then open the 500 — ran twice and held: backbone words first, then characters in two reviewed waves. Characters confirmed as the product’s heart (“i like the character ones the best”); the word entries’ rows now have char entries to link to. Timeline accepted as math, not hope: ~8–12 weeks of steady authoring runs to 4,000 entries; audio is days, not months. QA settled: no separate adversarial pass — author-side register checks + CI schema validation + Wedge spot reviews at batch gates.

Round 7 — uncertainty budget loosened. Supersedes Round 3’s zero budget in one direction: contested origins render when the contest helps encoding, presented as competing pictures (“the old graph carries two tellings”), never competing doubts; hedge vocabulary stays banned. 方 is the worked example (lashed boats vs plow, both told). The general decision test on student surfaces, in Wedge’s words: does it help the student.

Round 8 — the encoding-review amendment batch. From the three-document review, eight suggestions; outcomes:

• Lookup-first is the product identity. Definitions show immediately; encoding interactions (guess-commit-reveal, audio cloze) become an encode-mode renderer toggle, off by default — guessPrompt reserved in schema, authoring deferred. This tempers the deep-dive’s strongest recommendation against the product’s actual job.
• Drift badges adopted (the fence principle). Structured drift on sound-role composition rows: seven-group badge + which of initial/final/tone shifts. Piloted on 忙 ({m,w}), 幫 ({b,p,f}), 決 ({g,k,h}↔{j,q,x}). Converts sound components from trivia into prediction tools — the feature Wedge had never considered and liked most.
• Confusion sets are contract (失/矢, 衰/哀 — separating feature named), senseRef tags examples to tree nodes, per-reading trees for polyphones (行 converted; storage splits regardless of display toggles).
• Personal-context sentences are an overlay, never corpus. Resolves the generalizability concern: the published corpus stays universal; per-user reader-history sentences attach via the existing custom-layer precedence. The feature travels; the data doesn’t.

Schema bumped in place (entry@1: +guessPrompt, +drift, +senseRef, +meanings.byReading) — before the 500 are born into the old shape. Backfill across the 53 is the gate task before the 500 batch opens.

Ruled out

• Guess-first as the default page surface (it’s a lookup dictionary; encoding is the bonus).
• Personal video sentences inside the published corpus.
• Changes to story length and tree pruning — the review confirmed current practice (labeled temporary scaffolds; anchor senses only).

Outstanding

• Task #9 backfill (drift, senseRef, 便/擔 reading splits, lookalike sweep) → then the 500 opens, TOCFL backbone.
• Wedge pushing tsumugu-ed (53 entries + schema + CI) — in progress as this was written.
• Composer queue: display work order, voice for batches 2–4 (headword + example clips only), drift-badge chip rendering.
• Parked: encode-mode toggle design, audio hosting, license legal check (pre-P3).