[edkerk]

Moves develop to the RAVEN 3 YAML format that the release build writes, and makes the checks and the release accept it. Found in the release checks on #1123.

Why. Human-GEM up to 2.0.0 used RAVEN 2's writeYAMLmodel format: quoted strings, lists indented four spaces, integer charges. RAVEN 3 and raven-toolbox write a different format: quotes only where YAML needs them, two-space lists, every number as a float (charge: -1.0), and a subsystem as a list. v2.0.1 was already released in it, and the 2.1.0 build writes it again. Develop stayed in the RAVEN 2 format, and the checks only knew that one, so on the release branch:

• YAML lint failed on the list indentation;
• structureConsistencyTest read 0.0 as no charge (6334 findings instead of 57);
• the SBML round-trip lost every charge (a raven-toolbox reader bug, fixed in fix: read whole-number YAML charges as ints CompMetEng/raven-toolbox#164).

Changes

• model/Human-GEM.yml rewritten with raven-toolbox's write_yaml_model, as the release build writes it. No model content changes. Reading the file before and after gives identical metabolites, reactions, genes and compartments. The only differences are representation: a reaction's subsystem becomes a one-element list, and the metaData defaultLB/defaultUB become numbers.
• YAML lint: indentation: {spaces: consistent, indent-sequences: whatever}, accepting list items at their key's indentation.
• structureConsistencyTest: reads charges written as floats.
• qcModelChecks split-compartment check: accepts a subsystem written as a list. Without this, the artificial reactions were no longer skipped (37 instead of 26).
• increaseHumanGEMVersion.py build: stamps the release date. RAVEN 2 stamped the current date on every write, but the RAVEN 3 writer keeps the stored one, which is why v2.0.1 still carried 2026-03-26.
• release.yml: the sync-back keeps develop's Human-GEM.yml, since the build stamps the version into the release copy.

Tested on Vera, with the converted YAML:

• yamllint: 0 errors.
• cobra and raven-toolbox round-trips: lossless.
• SBML round-trip: models semantically equal, with raven-toolbox#164.
• Structure consistency: 57, as before.
• qcModelChecks: unchanged from develop (split compartments 26, growth 1.90).
• Metabolic tasks: all 57 essential and 21 verification tasks pass.

Needs CompMetEng/raven-toolbox#164 merged first, since CI installs raven-toolbox from its default branch; until then the SBML round-trip here will fail.

Also found: the RAVEN 3 writer does not quote YAML 1.1 booleans, so formula: NO (nitric oxide, MAM02609) is read as False by YAML 1.1 parsers such as PyYAML. cobra and RAVEN are not affected; yamllint reports it as a truthy warning. To be raised in raven-toolbox.

[github-actions]

Model quality report

⚠️ 6 pre-existing finding(s), no regressions vs develop. Non-blocking. Gates: duplicate keys 0 ✅, growth 1.9 ✅.

Section	Status
Model & network checks (20)	⚠️ 6 pre-existing
Model file & metabolic tasks (6)	✅ all pass
MEMOTE	✅ 63.8% (core subset)
Full report	model_qc_summary.md

Pre-existing:

• Reactions split across compartments: 26
• Reactions flagged by MACAW dead-end test: 1111
• Reactions flagged as MACAW duplicates: 340
• Mass-imbalanced reactions: 69
• Charge-imbalanced reactions: 172
• Structure vs formula/charge inconsistencies: 57

✅ unchanged · ✨ improved vs develop · ⚠️ pre-existing, non-blocking · ❌ regression

Full workflow run · this comment is edited as results come in