LLM-validated CCSS-math prerequisite graph + get_learning_path tool.
Full design: ~/.claude/plans/rippling-yawning-codd.md. State + resume notes:
memory project_prereq_graph_pilot.
prereq_candidates.py— Phase 0. Regenerates ~2,568 candidate prereq pairs from a scratch fork of prod (CCSS-math content, strictly-lower-grade, cross-domain, cosine≥0.45, grade-window≤3, top-6, MP excluded). Deterministic. Repoint DB path + output path to a fresh scratch dir before running.prereq_gate.py— Phase 1. Directional HARD/SOFT/NONE gate on Studio qwen2.5:72b, validated against the 40-pair gold set. The GOLD dict (38 hand-labeled pairs keyed by id-suffix) is the non-regenerable treasure — 20 HARD / 10 SOFT / 8 NONE, with MUST_NOT_HARD (spurious traps) and MUST_HARD (real cross-domain prereqs relate.py misses) as critical sets. Bar: ≥85% binary (HARD vs not) + 0 spurious accepted + 0 cross-domain missed.prereq_gold_sample.json— the 40 sampled candidates (texts) the gate reads.
Paths inside the scripts point at a now-defunct session scratchpad — repoint
them. Scratch DB is prereq_pilot.db (a .backup fork of prod).