Tsumugu: the grammar encoding layer opens (2026-06-13)

The next layer after vocabulary. Deep dive over ICS-CONDENSED, OUTLIER-CONDENSED, the encoding deep-dive memo, and the GSM2 dialogue corpus, to answer one question: how do you build grammar pages that encode rather than explain? The founding insight is that the existing system encodes objects — a character is form + meaning + sound, decomposed into functional components — but grammar is a procedure, a map from an intention in the head to a structured utterance, so it needs its own decomposition. A PRD is drafted with a fully-worked 把 example. Wedge ratified the model (“your analysis and assessment is correct”) and reshaped the Encode mode. This entry exists so the foundation isn’t lost before we ratify the rest one-by-one — it sets the contract for every grammar entry.

Decisions

Grammar gets its own functional decomposition — the four faces. Function · Form · Trigger · Boundary, mapped onto Outlier’s form/meaning/sound/empty so grammar inherits Outlier’s authoring discipline wholesale (typed roles, labeled exceptions, the fence principle, grounding). The novel face is Trigger: the cue in your own thinking that should make you reach for the pattern — grammar’s analog of Outlier’s “anchor to a word you already know,” here anchoring to an intention you already have.

Production pressure is a property of the pass, not the page. Wedge’s reframe, and the thing that broke the Refold-vs-ICS deadlock. One canonical, exhaustive, low-pressure rulebook, consumed through selectable modes. Default low-pressure; output opt-in; modes learner-selectable and system-suggestable by maturity. This reconciles his own Refold note (grammar = low-pressure noticing aid) with the brief’s “teach me to arrange my thoughts when speaking” without choosing a side.

The three modes — Encode is the heart. Prime = the prestudy primer, ~75s, zero output — ratified as-is (“prime is great”). Encode reshaped this session: not relentless guess-commit-reveal (it aggravates and overwhelms), but the “how to think about this grammar point / how to arrange it in your head” mode — Thought Order takes precedence, with the Form frame and the structural Story behind it; testing stays low-stakes. Produce = opt-in deep-processing / test-prep (construct, speak, record, compare) — kept, pending confirm.

Thought Order is the centerpiece. The intention → utterance walk, ending in a self-check gate — the operational form of ICS’s “verbalise step” (Summit 224), exploiting that Mandarin word order largely mirrors real-world/temporal sequence. The brief’s whole “arrange your thoughts when speaking” goal lives here, and it now leads the Encode mode.

Flagship worked example = 把. Built end-to-end, grounded against the Chinese Grammar Wiki (definite object; mandatory result complement; negation before 把; 把/被 as flipped spotlight). The anchor that ties the layers together: 把 = 手 (hand) + 巴, literally “to grasp” — the character’s meaning is the grammar’s function: grab the object, drag it to the front, report what became of it.

GSM grammar videos feed via Wedge’s notes, not transcription — only the dialogues are captured as text in the repo. Reuse over rebuild: entry@1 + a new kind: "pattern", the player, the grounding system, the activity-ladder modes. No parallel system.

Ruled out

• Production-first pages — the “production rulebook / translation manual” Refold explicitly warns against, and ICS’s no-forced-output-on-new-items rule.
• Guess-commit-reveal as the default interaction everywhere — downgraded to sparing, low-stakes use in Encode, per Wedge: don’t aggravate or overwhelm the learner.
• A single fixed point on the noticing↔production spectrum — replaced by selectable modes.
• Transcribing the GSM grammar videos now — notes instead.

Outstanding

Foundational decisions to ratify one-by-one (each sets the entry contract):

• Encode shape — confirm Thought-Order-first + Form + Story + low-stakes tests.
• Produce mode — keep as the opt-in output pass, slim it, or defer to v2.
• Encoding gate level — per-point recognition vs production graduation (lean: per-point gateLevel, default recognition).
• Granularity — 了₁ / 了₂ as separate entries vs one entry with branches (lean: split by function, polyphone-style).
• Inventory & ordering — master grammar-point list source (GSM sequence / TOCFL / Chinese Grammar Wiki / reader encounter order).
• File home for the PRD and the pattern entries; the form Wedge’s GSM grammar notes will take.

Next move: ratify the above, then lock entry@1 kind: "pattern", author the 把 JSON as the canonical seed entry, and render-check the three modes from one entry.

Update — the attack-vector pivot (later)

The first layout draft came back too rulebooky — a comprehensive explainer, “no different from Googling it,” with no mnemonic-first encoding. It broke Refold’s own rule (don’t be comprehensive, don’t be a rule-manual; lead with what’s noticed early). The fix, in Wedge’s word: an attack vector — the brain’s grappling hook that attaches straight to the concept. Fitting, since 把 is a grasping hand.

• Encoding hook leads; rulebook collapses to reference. Every grammar entry opens like a character entry — Form + Story — then folds the rules / examples / edges into expandable detail. Refold-light on top, full depth on tap. Dissolves the earlier rich-vs-light tension by making it a vertical hierarchy. (decided)
• Form is the visible composition, not a definition. form shows the part you can see in the character (把’s hand, 扌; 巴 carries the sound), never a gloss like “to grasp.” story is the hook: the hand grabs the thing and drags it to the front. “The hook” and “attack vector” stay internal terms, never shown to the learner. (corrected)
• “Name the attack vector” = mandatory first authoring step. No entry ships without its hook. The slot template (old “Form”) renamed Frame, demoted to reference. (decided)
• Open risk: abstract patterns (是…的, resultatives, topic-comment) have thin Form-meaning — their hooks must be invented (spatial schema / gesture). Hook quality is what makes or breaks the layer. (open)

Producer note: ICS / Outlier / GSM prose paraphrased, never republished. PRD: PRD-Grammar-Encoding-Dictionary-2026-06-13 (home TBD).