Five ideas, each independently gated

A GAT over a minimally-scoped Obligation node graph, with a 3-feature edge-attention scheme (modality compatibility, temporal overlap, legal-semantic similarity) and a weak-supervision plan distilling the existing rule-based conflict finder. The literature/patent search and GNN training remain open.

Extracts “N days/months after <trigger phrase>” conditional patterns — distinct from absolute-date extraction — and, given hand-provided anchor dates, auto-derives downstream dates. Validated against five hand-modelled scenarios. The full portfolio TRIGGERED_BY-graph + Monte-Carlo version needs KG-schema growth.

122 verified (anchor, positive, hard-negative) triplets built with named perturbations — modal-verb swap, negation, day-count/dollar/percentage shift — each confirmed to actually change legal effect. The contrastive embedding fine-tune is the GPU-blocked step.

A counterfactual fingerprint-delta attribution method plus a classical acceptance predictor. The predictor does not beat its majority baseline — a precise finding: unsigned delta magnitude can't encode edit direction, which needs idea #3's signed learned subspace. Blocked further on real redline history that doesn't exist yet.

Perturbs one legally-meaningful span at a time — a deontic marker, an entity, a defined term — and measures the shift in predicted-class probability. Found near-zero attribution to the deontic marker on the real gold set, independently corroborating a SHAP-based finding that the trained risk model learned superficial n-grams, not legal structure.