The vocabulary of the domain. These terms are used precisely and consistently in code, API and docs. Where the Dutch pack has its own word for something, it is given for reference only — the platform never uses it.
Pack — a complete training program for one skill, e.g. Dutch conversation B1→B2. A pack declares its own content language, its competency framework, its error categories and its section map. Everything skill-specific enters the platform through a pack; nothing about a skill is hardcoded in the runtime.
Method — a pack's declared teaching method: the lesson arc, its authoring rules, and its
per-topic sequencing notes. Where the ramp says how hard the next block should be, the method
says how it should be built. Both are carried into the brief verbatim and neither is interpreted.
See docs/guides/teaching-method.md.
Block (Dutch pack: "blok") — an ordered, themed batch of lessons within a pack, with a level, a grammar/skill focus and a milestone. Blocks are the unit of authoring and of adaptation: only the current block need exist, and the next one is written from evidence about the last.
Lesson (Dutch pack: "les") — one sitting. Ordered within its block. A lesson is an ordered array of typed sections.
Section — a typed piece of a lesson. The runtime knows nine kinds: text, rules, vocabulary,
questions, speak, write, listening, dictation, exercise. The viewer renders each kind its
own way; write sections are what produce a submission.
Drill item — one practisable atom, held in a deck separate from lesson prose so it can be
scheduled independently. Three kinds today: term (a word or phrase with its translation and an
example), word-order (a sentence, its correct chunk order, a translation and a grammar tip), and
mcq (a question, its options, its answer key, an explanation, and the error categories it tests).
Answer key — the correct options an mcq carries. Held on the server and revealed only after
the learner commits. Because the key is authored in advance, a wrong answer records an error-log
occurrence with no coach involved — see
ADR-0014.
Multiple response — an mcq whose key names more than one option. Scored all-or-nothing: every
correct option and no incorrect one. There is no partial credit and no override.
Sitting (a quiz session) — a fixed set of questions answered through in one go, in one of two
modes. practice grades each answer as it is given; exam withholds every verdict until the end.
Its score and per-category breakdown are derived on read, never stored.
Own word — a term item a learner added themselves rather than one that arrived with the pack.
It practises identically and lives in the same deck, but only its owner can see it and a republish of
the block never removes it. See ADR-0012.
Article — a piece of long-form reading loaded into one learner's library. Not part of a pack's curriculum: it is the learner's own domain brought into the pack, so nobody else ever sees it. An article is parallel text — one variant per language — and is identified by its slug, so re-loading it corrects it in place. See ADR-0017.
Variant — one language's rendering of an article: a language tag, a title, and markdown. Which variant a learner sees is decided by the interface language, falling back to the pack's content language. This is the only place in the product where the language switch changes material.
Label — a free string an article carries, chosen by whoever loaded it. The library filters and groups by labels; the runtime never interprets one, exactly as it never interprets a ramp's dials.
Content language — the language a pack's material is written in. Distinct from UI language, the language of the interface chrome. A Dutch pack renders as Dutch whether the interface is set to Dutch or English. See ADR-0005. The one narrowing of this is an article's variants, above.
Learner — a person working through a pack, identified by the sub claim of their
identity-service token. The learners record is a thin profile (display name, UI language); Skills
Coach stores no credentials.
Enrollment — a learner's position in a pack: current block, current lesson.
Drill state — per learner, per drill item: which stage they are on, their current streak, and whether the item is mastered. This is the spaced-repetition state.
Reading state — per learner, per article: that they marked it read, and when. Kept apart from the article for the reason drill state is kept apart from the drill item — content gets re-loaded, and a corrected translation must not arrive as something unread. Reversible: read is a filter the learner sets, never a measurement of them.
Stage — a direction of practice for one drill item. Stage 2 is gated behind mastering stage 1.
For term items: stage 1 is content-language → translation, stage 2 the reverse (which drills
spelling). For word-order items: stage 1 is the primary correct order, stage 2 an alternative valid
order of the same chunks.
Streak — consecutive correct answers at the current stage. Two clears the stage. Any wrong answer resets it to zero — a near-miss is not partial credit.
Attempt — an append-only record of one answer to one drill item: what was given, whether it was correct, and whether the learner overrode a rejection.
Submission — a learner's written answers for a lesson, awaiting correction. pending until an
external coach corrects it, then corrected.
Correction — what the external coach returns for a submission: per-item {original, corrected, categories, explanation}, a category tally, and advisory ratings. The runtime derives error-log
changes from it; the coach never writes counters directly.
Error category — a stable, pack-declared label for a kind of mistake (the Dutch pack uses
woordvolgorde-bijzin, perfectum/imperfectum, and so on; the AWS pack uses the exam guide's twenty
task statements). Categories are the join key between correction, drilling, quiz questions and
next-block generation, so they must stay stable across a pack's life. A category may declare a free
group label — an exam domain, say — which the viewer groups by and the runtime never interprets.
Error log (Dutch pack: "foutenlog") — per learner, per category: examples, first and last seen,
a count, and a status of new → recurring → improving → mastered. This is the memory that
makes adaptation possible.
Block review — the end-of-block assessment, carrying a next-block brief: what to re-drill, what to retire, and the theme and difficulty for the block after this one.
Brief — the assembled input for authoring the next block: current error-log state, position on
the competency ramp, the pack's method, and the program goal. Fetched from GET /coach/v1/blocks/:id/brief. The runtime
assembles it; an external coach consumes it.
Runtime — this system. Stores, grades deterministically, counts, progresses. Generates nothing.
External coach — whatever authors lessons and corrects free-form answers. Today a person driving
an LLM CLI against /coach/v1; the contract is designed so this can later be the runtime calling a
model API without the API changing.
Capability — an action Skills Coach permits, e.g. submission:write. Roles arrive on the token
from identity-service; Skills Coach owns the role → capability map and the enforcement. An unknown
role grants nothing.