Checkpoint 1 · Grammar Phase B (PLAN-177 / ADR-084 Amendment 2026-07-06)

Grammar concept curriculum: F10 is live, review the calls

The grammar rewrite from the 2026-07-03 pilot verdict shipped its first real family. This page walks what's live, shows a real card, lets you confirm the 8 defaults I picked where you hadn't ruled, and asks your call on 5 genuine forks.

19 families / 175 rule shells authored F10: 40 of 54 rules fully enriched 834/922 forms linked to rules (90.5%) Dealt + graded live on the carousel

What shipped

The concept curriculum is real, not a mockup

Re-told plainly — the mechanics, not the pitch.

New three-tier archetype. Family container (never graded) → rule lesson (the graded unit — "I can state what this rule does") → future practice-cloze (gated, comes later once a learner is ready — not built this phase).
Every rule lesson had to pass YOUR content bar before it could save — a machine validator, not a vibe check: a 10-year-old reading level (sentence-length caps enforced), no unglossed jargon ("word-piece" not "prefix"; any unavoidable term gets a chip-gloss on first use), a situational example first (never opens with the rule), front-loaded vocabulary, 3-6 paired examples in labeled groups, side-by-side English, and an explicit contrast with how English does it. That's the skeleton of Atmosumarto's teaching structure, not his prose — no verbatim copying.
Rules now link to your existing word-forms. A new field on the 922 form cards (form_rule_keys) stamps which rule(s) each form instantiates, from a committed, auditable pattern map — 834/922 = 90.5% linked. A rule's example forms are now a live query against real corpus data, not hand-picked examples.
Multilingual worked exactly as you predicted — lessons ride the same {Indonesian, English} translation-tracked field shape as the rest of the catalog. No new partition, no fork. Verified before any content was built, not assumed.

19 / 175

families / rule shells authored

40 / 54

F10 rules fully enriched

90.5%

forms linked to a rule (834/922)

$0.75

total Gemini spend, this phase

See it live

A real card: why some verbs take nothing, some ber-, some meN-

This is our own generated content — safe to show in full. Terminology bar: plain gloss, no jargon, side-by-side examples.

Rule lesson · F10 · A2

simple-vs-ber-vs-men

Plain-language gloss Some verbs need no word-piece. Some add ber- to the front. Some add meN- to the front.

IndonesianDia mandi setiap hari.

EnglishHe bathes every day.

Anak itu berlari cepat.

That child runs fast.

Saya menulis surat.

I am writing a letter.

Your architecture, confirmed

What you already ruled (conversational, 2026-07-06)

Landed in the ADR-084 Amendment as-stated — no defaults needed here.

Lessons teach concepts at the catalog level, per L1-L2 pair

Form-learning (CT3) stays practical per-sense reinforcement; the two layers cross-link via form_rule_keys.

Decided by Frank

Multilingual rides the existing t070 pattern

Verified live before building: every prose field is already a translation-tracked {ind, eng} shape. No new schema needed.

Decided by Frank

Enrichment adds concept-lesson generation + a rule↔form linking field

The link is a first-class DB column (form_rule_keys), stamped by a deterministic, auditable pattern map — not LLM guesswork.

Decided by Frank

Review — 8 items

Adopted defaults (AD1-AD8)

You hadn't ruled on these directly — I picked the option that best matched your stated direction. Confirm each, or flag one to revisit.

AD1 — 3-tier shape confirmed. Family container / rule lesson / gated future clozes — matches your v3 answer.

AD2 — overridden by your instruction. The rule↔form link is a first-class DB field (form_rule_keys), not pocket-only. member_card_ids was dropped from the contract.

AD3 — grammar track is opt-in, never default-on.

AD4 — dedicated rule renderer + graded-partial-reveal hint deferred to next UI phase. v1 lessons render through the existing card blocks (proven live above).

AD5 — audio autoplay OFF by registry default for rule cards. No example audio clips exist yet; flagged so manual-play survives when they do.

AD6 — lifecycle staged. Registry declares the gated tier-3 clozes, but v1 runs no-forms; the standard lifecycle + family-learning gate land together at stage-2 (that gate is unbuilt for vocabulary too). No clozes generated this phase.

AD7 — generation runs as a dedicated serial script, not a queue job kind. The daemon's 'rule' job type is deferred until the daemon is v3.6-live — it only runs headword jobs today.

AD8 — CEFR bands are PROVISIONAL. Stamped from your skeleton's level tags. A few worked examples in the rubric doc disagree (logged as regrade candidates below in FK-adjacent detail) — regrading just means editing the inventory and re-running, no hard-coded rubric.

Decisions needed — 5 items

True forks (FK1-FK5)

These are genuinely undecided. Pick an answer or leave a note — either moves the plan forward.

FK1 (= your Q4) — multi-level topics, like -nya spanning A2/B2/C1. Designed-for via a topic_key seam — a topic can split into level-tagged sibling cards with zero rework, whenever you say go. Nothing is split yet.

FK2 — the Sneddon completeness gap. Of 727 Sneddon sections, 608 are covered by the skeleton; 119 are not — rolled up: 31 more affix sections, 21 adjuncts/adverbs, 13 pronouns, 12 word-forming (full reduplication etc.), 12 numbers, 7 nouns, 6 mood, 5 complex sentences, plus smaller categories. Per your "Sneddon = completeness" instruction — tell me which deserve skeleton nodes and I'll add them (they'd author as shells, enrich in a later pass).

FK3 — the F10 tail. 14 rules are stuck: 8 enriched in Indonesian but the English translation can't hit the 25-word sentence cap after retries (kept honest — no half-translated lies shipped), and 6 can't pass the reading-level/example validators after 3 attempts (they stay skeleton, fully resumable). These are the genuinely hard rules; a targeted re-pass, or a slightly relaxed cap FOR TRANSLATIONS ONLY, would recover most of them. Your call on whether that's worth a follow-up.

FK4 — grammar-vocab terminology cards. Every lesson already captured terms_used + a glossary restatement (the back-of-book glossary shape). Building the terminology itself as concept cards — so lessons LINK to "what a prefix is" instead of re-glossing it every time — is a small follow-up plan whenever you want it.

FK5 — skeleton tally mismatch. Your skeleton doc's own count says 126 nodes but it actually contains 179 usable ones (the F10 breakdown also undercounts). The inventory records reality and is the working source of truth; the raw doc itself could use a tally fix, but I didn't touch it — raw/ is your inbox, not mine to edit.

Process notes

Two things I did that you should know, and how this ran

Model switch mid-build. Fable hit its weekly cap during the F10 straggler sweep; you authorized Opus for the cleanup/wiring remainder. Content generation (Gemini + Claude-CLI translation) is model-independent of orchestration, so lesson quality wasn't affected — Opus only drove the wiring and verification.
One dev-profile counter reset. To demonstrate a fresh drip release today, I reset the Grammar profile's daily drip counter (it resets itself at date-rollover anyway; dev profile, not user/FSRS data). That's the only data touch outside the pipeline.
ADR-084 got a dated amendment (per your prefer-amending-master-docs rule — no side ADR); peer review caught a would-be blocker before build (rule cards would have shown their family's title, the vows-build bug class, caught by a new write-path watch-for); your five mid-build directive batches (readability/terminology, Atmosumarto style, CEFR, rubric, variant-readiness) were folded into the generator before the full run; the enrichment builder read all 29 enrichment files (~16,500 lines) end-to-end first, per your requirement. Two validator fixes during the sweep were honest metric corrections (circumfix notation, a colon-label artifact) — the reading-level thresholds themselves were never relaxed.

Grammar concept curriculum: F10 is live, review the calls

Where this stands

The concept curriculum is real, not a mockup

A real card: why some verbs take nothing, some ber-, some meN-

What you already ruled (conversational, 2026-07-06)

Adopted defaults (AD1-AD8)

True forks (FK1-FK5)

Two things I did that you should know, and how this ran