Rule-based grammar findings

Where the project stands on harper-core (assessment spec 5.3, context pack 5.7).

The seam

curriculum::validate::GrammarCheck has one method, findings(&str) -> Vec<String>. Three places use it, and each takes it as an Option:

Place	With no checker
Rubric cross-check X5 (tutor-engine, ScorerEnv.grammar)	X5 is not evaluated. No alarm, no change of confidence.
Writing workshop, first layer (WorkshopEnv.grammar)	DraftSubmission.rule_checked is false and the list is empty. The UI must say "not checked", never "no findings".
Content warning W03 (UnitOptions.grammar_check)	The rule is reported as skipped.

There is no stand-in that returns an empty list. A missing checker is a missing checker.

Convention: a finding about spelling starts with Spelling. The rubric scorer leaves those out of the X5 count for a voice response, because a transcript's spelling is the recogniser's.

Status: linked (owner decision 2026-10-10, ADR-057)

Update 2026-10-10. The owner accepted the harper re-link. The tree now links harper-core 0.68.0 (default features off, plus harper's concurrent feature so the checker is Send + Sync):

Where	What	Spelling
assessment-engine	default-on grammar feature; GrammarChecker (the concrete checker, &self + internal lock)	per call
tutor-engine	default-on grammar feature; HarperCheck implements the seam for ScorerEnv/WorkshopEnv	on
content-cli	its own HarperCheck for UnitOptions.grammar_check (W03)	off on purpose
app-core	one HarperCheck built in Catalogs::load (one dictionary load per process), shared by the unit runs and the workshop	on
tutor-cli	builds one for unit run and unit score-pending	on

Version. The pre-merge line linked 2.11.0, but 2.11.0 (and 0.69+) cannot resolve in this tree: harper-pos-utils → burn → tracel-llvm-bundler → liblzma-sys links the native lzma, which conflicts with sherpa-onnx-sys → zip → xz2 → lzma-sys. The pre-merge tree had no speech/sherpa, so the conflict never appeared there. 0.68.0 is the newest version that resolves next to sherpa (checked empirically 2026-10-10: 0.55–0.68 resolve; 0.69, 0.70 and 0.72 fail). It still pulls burn 0.18.0 (272 new lock entries) but no second native lzma link.

Licenses. harper-core, harper-brill and harper-pos-utils say Apache-2.0; ammonia says MIT OR Apache-2.0; cssparser, dtoa-short and colored say MPL-2.0 (file-level copyleft, allowed for unmodified crates.io dependencies by per-crate exceptions in deny.toml and the register row). cargo deny check licenses passes.

The split, and why it stays split. W03 runs with spelling off: units are full of names the dictionary does not know, and W03 is about grammar. The workshop keeps spelling on: a typed draft is the learner's own spelling. The rubric's X5 drops spelling findings for a voice response (the transcript's spelling is the recogniser's) and keeps them for typed text. Do not "unify" the three.

What the linked checker actually finds (measured 2026-10-10, 0.68.0)

Run against the same sentences the 0.54.0 five-sentence check used:

Sentence	Findings
I has a book.	none
She go to school every day.	none
This is an test.	Miscellaneous: Incorrect indefinite article.
Last weekend I has a picnic with my family near the lake.	none
teh cat sat on teh mat.	spelling, capitalisation, typo findings
I recieve a letter.	Spelling: Did you mean to spell 'recieve' this way?

The agreement gap persists in 0.68.0. Subject-verb agreement is the most common error of the target learners, and the checker does not catch it, so X5 built on this version is a weak alarm and the workshop's first layer is a partial one. This is recorded, not hidden: the tests assert what the checker does find (the article mistake, spelling) and never assert that a clean result means correct grammar. Whether to layer a rule-based agreement check of the project's own on top is an open question for the owner.

History

• 2026-10-06: read the licenses, built the adapter against 0.54.0 outside the workspace, measured the five sentences.
• 2026-10-07: the owner approved harper-core; the pre-merge line linked 2.11.0 with spelling off for W03.
• 2026-10-08: the merge kept origin's tree, dropping the linking (ADR-055).
• 2026-10-10: the owner accepted the re-link (ADR-057). Linked 0.68.0, the newest version that resolves next to sherpa; all four call sites wired; the agreement gap re-measured and recorded.