Methods

A Python-native MBSE metamodel that couples SysML v2 semantics with Reinforcement Learning with Verifiable Rewards for agent training.

Python-native MBSE metamodel coupling SysML v2 semantics with RLVR. Four sub-models — OSM (BDD scenarios), SSM (typed system-state graph with declarative constraints), UOM (Set-of-Mark / Trace-of-Mark UI grounding), VM (verifier and reward) — share a URI-based traceability spine, with API ground truth as oracle and UI traces as the training target.

Multi-source belief-function combination over a restricted frame of discernment, yielding `[Bel, Pl]` intervals at every hierarchy level.

Six evidence sources — ColBERT MaxSim retrieval, gradient-boosted prediction, regex patterns, name matching, ModernBERT-NHSVM, and an LLM convergence agent — emit mass functions over a restricted focal set derived from the taxonomy hierarchy; Dempster's rule combines them into belief intervals `[Bel(A), Pl(A)]` at every level, with the interval width `Pl − Bel` quantifying epistemic uncertainty and the Dempster conflict `K` diagnosing source disagreement that point estimates suppress.

Classification by ColBERT MaxSim retrieval against an iteratively-optimized augmented controlled vocabulary, with per-text-feature analysis driving discriminative power.

Each ontology code is represented by an augmented definitional text — label, LLM-generated description, prototype values, name hints, value patterns, parent path, mnemonic — encoded once into a ColBERTv2 multi-vector and stored in Qdrant. Target items are classified by MaxSim retrieval against the vocabulary collection; a reflective evolution loop iteratively refines the vocabulary representation using per-feature attribution on both sides to drive discriminative power.

A BFO/CCO-grounded OWL ontology entails its SKOS vocabulary and SHACL shapes; the relational schema is a projection of that certified artifact, versioned so each commit is one coherent snapshot.

Ontology-grounded derivation in which HermiT-certified OWL is the source of truth, SKOS and SHACL are entailed views of the same axioms, and loadable SQL is lowered from the certified ontology. Every git commit is a content-hashed snapshot of the whole chain, enabling blind-release evaluation.

Hierarchy-aware multiclass SVM with a Crammer–Singer joint margin over ModernBERT mean-pool embeddings, post-hoc temperature-calibrated.

Normalized Hierarchical SVM (NHSVM; Choi, Chung & Hewitt 2015) head trained with the Crammer & Singer (2001) joint multiclass objective over 768-dim ModernBERT-base embeddings, with a structured tree-distance margin and a post-hoc softmax temperature fitted against a held-out reference slice for calibration.

A paired dataset and pipeline capturing every prompt, intermediate topic-clustering state, and verifier component score for constrained-decode-generated ontology compositions.

Per-sample preservation of the full input-intermediate-output trace — prompt, raw completion, parsed composition, verbalisations, KMeans topic centroids over a fixed reference clustering, and four-component verifier scores — for rejection-sampled ontology compositions, enabling exact verifier re-derivation, alternative scoring, and RLVR self-distillation against a stable embedding-space anchor.

Render a Metaflow DAG as a NiFi processor graph and route the flow's OTel spans into per-step processors on that same canvas.

Project Metaflow flow classes onto the NiFi canvas — one processor per step, edges following self.next() — and stand up a sibling ListenOTLP graph that routes spans emitted at runtime back to per-step LogAttribute processors. The canvas becomes both DAG diagram and live telemetry surface against a shared step-name key.

A failure-mode catalog gated by Risk Priority Number, escalating unresolved incidents to a coding agent that climbs a five-order RCA ladder terminating at CP-SAT constraint vocabulary.

Couples a quantitative FMEA catalog (S × O × D) to an Agent Client Protocol escalation path and a five-order root-cause abstraction ladder. Incidents the runtime cannot close are classified as OPERATIONAL or ARCHITECTURAL; the latter open GitHub issues that link symptom to a violated scheduler constraint.