p-061 — Synthetic field-coverage corpus

Status: completed. The generator, the 319 probes, the generated README with its predicted verdict per tag, and the --check guard in CI all shipped.

The corpus is the instrument, not the reading. Committing a measured verdict beside the predicted one — the first full run scored 2470 source fields down to 1891 reconstructed with 0 fabricated — would be a separate generated artifact and needs its own plan; see "Out of scope" below for why the measurement deliberately isn't pinned in the test suite.

Problem

Round-trip assumptions have been checked one field at a time, reactively: 655 was 100% lost, then 336, 370, 511 and 518. Each was found by noticing a gap in a real-record run and tracing it by hand. There is no systematic instrument that answers "for every MARC field the pipeline claims to handle, what actually survives?"

The curated fixtures are real records: excellent for fidelity questions, useless for coverage — they exercise whatever fields those particular Helmet records happen to carry, which is a small and uneven slice.

Design

Two synthetic records per MARC tag. <tag>.xml carries only the tag's primary subfield; 1<tag>.xml carries the maximal union of subfields both directions know about. The pair is the diagnostic, and it earned its keep immediately: a single maximal-only design reported MARC 020 as lost, when in fact 020 $a alone round-trips perfectly and it is the spurious $z/$q/$0 company that breaks it. Without the minimal variant, eight identifier fields (010, 017, 020, 025, 027, 030, 032, 088) would have been filed as pipeline gaps rather than subfield-combination breaks.

Isolated content. A record carries the leader, 001, 008, a 245 (so Boundary-1 minimum content passes) and only the tag under test. If a tag is lost, exactly one record shows it — no cross-field interference, no invalid field combinations.

Filename and 001 always match (655.xml carries 001 = 655; the maximal 1655.xml carries 001 = 1655), which keeps both inside the bare-digits filename pattern Boundary 1 accepts and makes roundtrip-eval's 001 pairing self-documenting.

Driven from the two existing coverage sources, so the corpus tracks the code rather than rotting beside it:

Source	Contributes
parse_xslt_corpus()	tags the XSLT reads, their subfield codes, indicator tests
MARC_EMIT_REGISTRY	tags the reverse converter emits, their subfields

Subfield values are greppable sentinels (Z655a) except where a value must be syntactically valid — vocabulary codes, authority URIs, relator codes, dates — which come from a curated table.

The generated README.md is the point of the exercise: per tag it records the subfields emitted and the predicted verdict from cross_check (round-trippable / forward-only / reverse-only / handled by neither). A run of the corpus then shows where reality departs from the prediction.

Trade-offs on the record

• Isolation over realism. One tag per record cannot catch interaction bugs (a 245 that changes how 246 converts). It is the wrong instrument for that and the right one for coverage. Interaction cases stay with the curated fixtures.
• $6 and $8 are excluded. $6 is the 880 alternate-script linkage and is meaningless without a paired 880; emitting it bare would produce a dangling link. The non-latin-translations fixtures already cover 880.
• Generated, with a --check guard, matching the three mapping docs. A submodule bump that changes the XSLT's field surface fails CI until the corpus is regenerated.
• Not a validity claim. These records are synthetic and minimal; they are not cataloguing examples and should never be mistaken for reference data. The README says so at the top.

Measuring it honestly

Two axes, not one. "Tag absent from the reconstruction" conflates three different things, and conflating them overstates the damage:

Outcome	Meaning
values kept, same tag	round-trips
values kept, retagged	data survives under a different MARC tag — a 655 with subdivisions returns as 650, because marc2bibframe2 renders it as a complex subject typed bffi:Topic
values lost	genuinely gone

The first pass at this analysis reported "59 asymmetric tags fully lost" by checking tag presence alone. That number was wrong: 30 of those tags keep their values under a different tag.

Out of scope

• Asserting on the corpus in unit tests. It is a diagnostic instrument the operator runs, not a fixture the suite pins — pinning it would freeze the very drift it exists to surface.