websemantic 0.2.6

Licence

• MIT from this version on. Versions 0.2.0 to 0.2.5 remain under PolyForm Noncommercial 1.0.0. The pinned scientific codes keep their own licences (TLS and LQL-Equiv: MIT; pyrcel: BSD-3-Clause).

Since 0.2.3

• SHACL checks of returned scenarios: grouped values (objects, lists, dotted and indexed paths) are flattened before mapping, without completing any value; unit tests added.
• Sensitivity of the contract effect to unparseable answers (evaluation/e1/parse_sensitivity.py).
• Spoken-request case: the stored first answers are checked by the Python validator and the SHACL rules; the specified follow-up is archived.
• Ontology 0.2.6: version-independent citation, MIT licence; FOOPS! 1.0 on 24 checks.
• Package published on PyPI (pip install websemantic); SECURITY.md and CONTRIBUTING.md; howfairis report archived.
• 536 tests pass.

Concept DOI (all versions): https://doi.org/10.5281/zenodo.23238902

websemantic 0.2.3

Persistent ontology identifiers

• Namespace https://w3id.org/websemantic/ns# (w3id.org); ontology IRI https://w3id.org/websemantic/ns; version IRIs https://w3id.org/websemantic/ns/core/X.Y.Z resolve to the tagged files.
• Preferred prefix wsem (registered in prefix.cc), logo, catalogue annotation, DOI and definitions for every core term.
• FOOPS! 1.0 on 24 checks at the ontology IRI (evaluation/e1/e1-foops-w3id-023.json); earlier reports kept for comparison.

Evaluation

• Preregistered controls: ontology length (L2/L3), generic clarification policy (G), held-out extension requests.
• Analyses: hallucination indicators, sensitivity of the pooled contract effect, SHACL range and relational checks on returned scenarios, interim analysis of running campaigns, spoken-request case: first answers and calculation from a specified follow-up (validation checks archived from v0.2.4).
• Full-precision p-values before Holm correction.
• ToolRosella baseline scripts (tool code not redistributed) and blind human-evaluation sample builder.

Maintenance

• Semantic packs and agent file manifests regenerated; 528 tests pass.

Concept DOI (all versions): https://doi.org/10.5281/zenodo.23238902

websemantic 0.2.0

One semantic source, three consumers: each scientific software is described once (descriptor) and compiled into an LLM contract, an ontology with SHACL shapes and FAIR metadata. Unchanged backends at pinned revisions: TLS (748e053), LQL-Equiv-web (dfc9a33), pyrcel 2.0.0 (977e909).

• Semantic packs for the three codes (variables, units with QUDT IRIs, qualified outputs, ontology, shapes, CodeMeta, FAIR statement).
• LQL-Equiv: schedule comparison (dominance criterion) and anatomical groups; qualitative dose wording is always asked.
• Evaluation harness (frozen contexts, every attempt logged), request-level analysis, qualifier sensitivity, HPC orchestration.
• Ontology metadata: FOOPS! score 0.87.

The reserved corpus, references and raw model answers are not part of this release.