External corpus runs — am::cadl::parse_definition

Not normative, and not a gate. This is the archetype half of tasks.md's "run an external corpus" item; json_corpus.md is the other half, canonical JSON compositions against serde_json → Composition → Validate. This file records what happened when archetypes nobody in this repository wrote were run through the definition reader: how many parsed, how many were refused and under which stated reason, and which refusals turned out to be this crate's defects rather than the corpus's. It exists because a conformance claim graded only by its own tests is the weakest kind (tasks.md P1, from the FerroEHR thread's post #15), and because every finding in audit.md was found by running something. The runner is openehr/tests/adl_corpus.rs, #[ignore]d, because the corpus is not in this tree (§Licence).

Read the counts the way the runner prints them. Parsed means the definition section produced a CComplexObject; it does not mean the archetype is valid, and it says nothing about the header, terminology, or rules sections, which the runner does not read. Refused means parse_definition returned a CadlError with a stated reason; the categories below are those reasons with per-file tokens stripped. A refusal is a pass for K15.6 (refuse by name) only when the name is right — the run's first finding (A-70) was a refusal under the wrong name.

How to run it#

# The corpus is read where it is. Never copy it into this tree (§Licence).
git clone --depth 1 https://github.com/openEHR/adl-archetypes /some/where
git -C /some/where rev-parse HEAD                 # record this with the run
cd openehr
OPENEHR_ADL_CORPUS=/some/where \
OPENEHR_ADL_CORPUS_REPORT=/some/where/report.tsv \
  cargo test --test adl_corpus -- --ignored --nocapture

The TSV has one row per file: extension, path, utf8/not-utf8, verdict, and for a refusal the category, byte offset, and raw reason.

Licence#

openEHR/adl-archetypes carries no licence file as of the commit below. Its archetypes are therefore read from a checkout the runner is pointed at, never vendored, never committed as fixtures, and never quoted here beyond a file name and the single line that reproduces a finding. That is also why the regression job tasks.md asks for — "fails when a previously-parsing file stops parsing" — does not exist yet: it needs a corpus this tree may carry.

Run 1 — 2026-09-03#

• Corpus: openEHR/adl-archetypes at 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: the tree at the commit that adds this file (A-70 included).
• Files: 1,379 .adls (ADL 2), 593 .adl (ADL 1.4).

Totals#

Before A-70 the same run parsed 178 .adls files and reported 4 with no definition section. The one that remains is ADL2-reference/validity/basics/openEHR-TEST_PKG-ENTRY.FAIL_definition_missing.v1.0.0.adls, which is meant to have none. The other three were Latin-1 files (a © in the description) that read_to_string had failed on — a runner defect, fixed by decoding lossily and counting the encoding separately. ADL 2 is UTF-8 by specification, so that column is a corpus fact worth keeping, not a parser verdict.

The .adl column is expected to be low: ADL 1.4's cADL differs from ADL 2's in ways this parser refuses by name (at-coded nodes are accepted, matches {*} and id-less objects are not), and the reader here is an ADL 2 reader that also happens to read the ADL 1.4 subset it shares (K15.5).

Refusals by stated reason, .adls#

Refusals by stated reason, .adl#

Findings this run produced#

• A-70 (fixed the same day): a differential-form attribute (/data/events cardinality matches {2..8; ordered}, openEHR-EHR-OBSERVATION.redefine_cardinality.v1.0.0.adls) mis-parsed into three attributes and was refused as VOKU (got ""). The category disappeared entirely after the fix — the number that says the diagnosis was complete — and parsed .adls went from 178 to 206.

Candidates — recorded, not yet findings#

Each needs a test written and the grammar re-read before it gets an A- number (W0.19, specification first; W0.3, nothing claimed unverified).

1. DATE_CONSTRAINT_PATTERN refused under the wrong name. date matches {yyyy-mm-dd} is c_date: ( DATE_CONSTRAINT_PATTERN | date_value | … ) in cadl2_primitives.g4, and the same for time and date-time patterns. am::cadl has no arm for a pattern, reads yyyy-mm-dd as an ISO 8601 literal, and refuses the d. Twenty-four files, all ISO 13606. The fix is a CPrimitive pattern form for C_DATE/C_TIME/ C_DATE_TIME (A-63 modelled the value forms only) — a real construct, so the model changes first, then the parser.
2. primitive_kind matches the type name case-insensitively. Read while examining the same files: an RM class spelled DATE (ISO 13606 has one) would be taken for the Date primitive. Not yet reproduced by a test; the corpus files above are refused earlier, on the pattern, so the run does not show it.
3. Refusal names that are correct but unhelpful. CIMI is not a valid id-, at-, or ac-code for an id-less use_archetype, and expected an RM type name, found * for ADL 1.4's matches {*}. Both are right to refuse; both could say what the writer actually did. K15.6 asks for a name, and these are names; whether they are good enough is a judgement for the next run to make with the ADL 1.4 reader in view.

Decisions the run asks for#

• effective_occurrences. 1,197 of 1,739 refusals across both extensions are the parser declining to infer omitted occurrences on a non-root node. AOM2 states the rule (C_OBJECT.effective_occurrences: lower bound 0; upper bound the owning attribute's cardinality upper bound if it is a container, else the RM's own multiplicity for the attribute, i.e. 1). The parser already knows which of the two cases it is in — a cardinality clause is syntactic. Implementing the rule is a K15.x requirement in 15-archetypes.md first; what it changes is that CObject::occurrences() would return an inferred interval a reader cannot tell from a stated one unless the type records which it was. That distinction is the whole decision, and it is the matrix's to make, not this file's.

Run 2 — 2026-09-03, after A-71#

• Corpus: unchanged, 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: run 1's tree plus A-71 (K15.32): an omitted occurrences is carried unstated and inferred from its owning attribute, not refused.

Totals#

The .adls count went from 15% to 56% on one change, which is what run 1's "two thirds of every refusal" predicted. The .adl count barely moved: ADL 1.4 files are refused earlier, on syntax an ADL 2 reader correctly does not accept (matches {*}, id-less objects, 0|[local::…] ordinals).

Refusals by stated reason, .adls#

Refusals by stated reason, .adl#

Findings this run produced#

• A-71 (fixed, breaking): the model itself could not carry an unstated occurrences; K15.32 written first, then Option<MultiplicityInterval> on every C_OBJECT type, effective_occurrences, and the parser carrying rather than refusing. The test fixture that had asserted a named refusal of the real openEHR-EHR-CLUSTER.device.v1.0.0 definition since A-62 now asserts the parse.

Candidates — added by this run#

4. An unwrapped interval's kind is decided by the grammar, not unknowable. A-67 refused integer_attr3 matches {|0..100|; 10} on the ground that C_INTEGER and C_REAL "cannot be told apart without a wrapping rm_type_id". odin_values.g4 says otherwise: integer_interval_value is built from INTEGER tokens (DIGIT+) and real_interval_value from REAL tokens (DIGIT+ '.' DIGIT+, base_lexer.g4), so |0..100| is a C_INTEGER and |0.0..100.0| is a C_REAL — the same lexical distinction this parser already applies to an unwrapped single value (Token::Integer vs Token::Real). 184 files across both extensions, the largest refusal left, and 120 of them the clinical corpus. The fix is to read the first bound's token kind and dispatch; the refusal should survive only for a bound mixing the two (|0..100.0|), which the grammar also refuses.
5. Unwrapped temporal literals. duration_attr1 matches {PT0S} and the date-time pattern above reach c_objects as a Word, are taken for an RM type name, and are refused as "expected [". c_inline_primitive_object admits every primitive kind, and ISO8601_DURATION, ISO8601_DATE, ISO8601_DATE_TIME, and the *_CONSTRAINT_PATTERN tokens each have a lexical shape (base_lexer.g4) this parser's read_iso8601 already recognises inside a typed primitive. 36 files, with candidate 1's patterns in the same place.

Decisions the run asks for#

• Reference Model multiplicity. A-71's residual and the 11-file row above are the same fact: without a table of which RM attributes are containers, items matches { with no cardinality clause is single-valued here and 0..* in AOM2. The table is small (the RM's List/Set-typed attributes) and this crate already has the RM in rm::; deriving it there, once, is the lib:A-33 shape — one rule, one home — rather than a second table in am::cadl. A requirement first.

Run 3 — 2026-09-03, after A-72, A-73, A-74#

• Corpus: unchanged, 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: run 2's tree plus A-72 (an unwrapped interval's kind decided by its first bound's token), A-73 (allow_archetype … closed parsed), and A-74 (the relop interval spelling |>=0.0|), the last found by the intermediate run between the first two and this one.

Totals#

By directory, the clinical corpora now lead: Reference/CKM_2013_12_09 237 of 322 parsed, Reference/Nehta_2014_04_25 132 of 164. The reference suite (ADL2-reference) is 219, a third of it the validity directory, where a refusal is the expected answer for the invalid twins.

Refusals by stated reason, .adls#

The .adl column is unchanged in shape: 471 id-less objects and matches {*}, the rest stated limitations.

Findings this run produced#

• A-72 (fixed): A-67's "cannot be told apart" was wrong; odin_values.g4 decides by token. 184 files.
• A-73 (fixed): the closed slot stayed refused after A-71 removed the reason. 13 files.
• A-74 (fixed): the relop spelling failed on its =, not by name. 106 files, surfaced only once A-72 let intervals reach the reader — a refusal can hide another, which is why runs are recorded one at a time.

Candidates — status#

1. DATE_CONSTRAINT_PATTERN (and the date-time form): open, 35 files.
2. primitive_kind case-insensitivity: open, still not reproduced.
3. Refusal names that are correct but unhelpful: open.
4. Unwrapped interval kind: closed, A-72.
5. Unwrapped temporal literals: open, 32 files; shares candidate 1's fix (a CPrimitive pattern form and unwrapped temporal dispatch by lexical shape).
6. New — the +/- interval spelling is refused by name and no corpus file uses it; recorded so its absence from the numbers is not read as support.

Decisions the run asks for#

• Reference Model multiplicity (run 2, restated): 28 files now, 19 of them clinical, and the count will keep rising as other refusals fall. The decision is in plan.md; a requirement first.

Run 4 — 2026-09-04, after A-75#

• Corpus: unchanged, 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: run 3's tree plus A-75: temporal *_CONSTRAINT_PATTERNs read for all four kinds, wrapped or unwrapped, and the four temporal kinds reachable unwrapped (c_inline_primitive_object's own full grammar, not the five-kind subset this parser used to admit).

Totals#

The d is not a valid ISO8601_DATE category (14 files, all Reference/ISO_13606) and its duration sibling (1 file) both dropped — but not to zero, which is what led straight to A-76: one file still failed the same way, for a different reason A-75 does not touch.

Refusals by stated reason, .adls, that changed#

Run 5 — 2026-09-04, after A-76#

• Corpus: unchanged, 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: run 4's tree plus A-76: primitive_kind matches exactly, not case-insensitively.

Totals#

The ISO8601_DATE/ISO8601_DURATION categories are gone entirely. One category rose in their place: children require more occurrences than the cardinality permits went from 1 to 8 files (7 in .adls, 7 in .adl, all Reference/ISO_13606) — not a regression, but the files parsing further than before and meeting a real VACMCU check the earlier misparse never let them reach. Not yet examined for whether the archetype or this parser's cardinality inference is at fault.

Findings runs 4 and 5 produced#

• A-75 (fixed): temporal patterns and unwrapped temporal literals, corpus run 1's candidates 1 and 5, closed together.
• A-76 (fixed): primitive_kind's case-insensitive match, corpus run 1's candidate 2 — recorded as "not yet reproduced by a test" there, and reproduced for real chasing A-75's own residual refusal.

Candidates — status after run 5#

1. DATE_CONSTRAINT_PATTERN: closed, A-75.
2. primitive_kind case-insensitivity: closed, A-76.
3. Refusal names that are correct but unhelpful: open.
4. Unwrapped interval kind: closed, A-72.
5. Unwrapped temporal literals: closed, A-75.
6. The +/- interval spelling: open, refused by name, still unused in the corpus.
7. New — a negative unwrapped integer in a C_ATTRIBUTE_TUPLE row ({-3}), 2 files, expected a primitive value, found-``. Not yet examined.
8. New — a VACMCU cardinality violation surfaced by A-76, 8 files, all ISO 13606. Not yet examined; may be the corpus's own defect, not this parser's.

Decisions the run asks for#

• Reference Model multiplicity (runs 2–3, restated): still open, plan.md.

Run 6 — 2026-09-04, after A-77#

• Corpus: unchanged, 093c77ea003742b9540e3dd377d615e2b26f2996.
• Crate: run 5's tree plus A-77: a negative unwrapped number ({-3}) is dispatched, wrapped or unwrapped.

Totals#

The expected a primitive value, found `-`` category is gone — both files that produced it (candidate 7) now parse.

Candidates — status after run 6#

1. DATE_CONSTRAINT_PATTERN: closed, A-75.
2. primitive_kind case-insensitivity: closed, A-76.
3. Refusal names that are correct but unhelpful: open.
4. Unwrapped interval kind: closed, A-72.
5. Unwrapped temporal literals: closed, A-75.
6. The +/- interval spelling: open, still unused in the corpus.
7. A negative unwrapped integer in a tuple row: closed, A-77.
8. A VACMCU cardinality violation surfaced by A-76, 8 files, all ISO 13606: examined and closed, no code change — the archetype's own inconsistency, not this parser's. Seven of the eight are Reference/ISO_13606/Spanish_MOH COMPOSITIONs that share one boilerplate SECTION: members cardinality matches {0..1; unordered; unique} matches { allow_archetype ENTRY[…] occurrences matches {1} … allow_archetype ENTRY[…] occurrences matches {1} … } — two mandatory children under an attribute whose cardinality permits at most one, contradictory on its own terms and correctly refused (CAttribute::container's own VACMCU check, A-71's residual documentation). Present verbatim in every file examined, so it reads as a systematic artefact of whatever produced these from EN13606, not a one-off. The eighth, ADL2-reference/validity/structure/ openEHR-EHR-OBSERVATION.WACMCL_container_items_out_of_bounds.v1.0.0.adls, is the reference suite's own deliberately invalid fixture — its name says so, and refusing it is the correct answer, the same as every other validity directory file this corpus contains.

Decisions the run asks for#

• Reference Model multiplicity (runs 2–3, restated): still open, plan.md.

Trademarks#

openEHR® is the registered trademark of the openEHR Foundation and is used with the permission of openEHR International. Use of the trademark does not constitute endorsement of this product by openEHR International or openEHR Foundation.