specs/017 + 018: loop pool (refuted) and derived-edge welds (generator fix adopted; held-out +173)

[adamjohnwright]

Loops: the pool (specs/017, refuted) and the derived-edge welds (specs/018, generator fix adopted)

Replaces #56 and #57, which GitHub closed when their base branch (#54) was merged and deleted. Same commits, rebased onto main by merge; every commit is in order (spec → plan → pre-registration → implementation → scoring → review errata).

specs/017 — loop as a conserved pool: recorded negative result

DS_SCC_METHOD=pool | pool_parity | pool_all, default off and bit-identical. Pre-registered, measured on the production catalog on both axes, refuted by its own rule: pool −53 held-out (it rails small loops the fixed point held still); pool_all −326 held-out, −17 experimental, AKT cases 100 → 2. Traced mechanism: product-over-entries is AND applied to alternative routes (RAS isoforms in one GTPase cycle). Two adversarial reviews folded in (baseline error corrected before scoring; pool ≡ pool_parity on this catalog; entry-test noise fixed; mid-run edit shown value-neutral by a clean re-run). test/test_loop_pool.jl 81.

specs/018 — where our loops come from

Diff against MP-BioPath's hand-curated networks (bipartite join on dbIds): they keep the reaction backbone inside our cycles (48–64% present) and never carry our derived classes. Census: removing assembly + depletion edges takes TP53's component 836 → 34, DSB 1,127 → 126. Cause (Adam's diagnosis): boundary expansion reused a leaf node the root complex itself reaches downstream — 1,994 of 2,077 cycle-carrying assembly edges. Fixed in the generator (reactome/logic-network-generator#91, targeted after review: refuse only downstream nodes; edge multisets identical to before in all 92 pathways, +336 fresh leaf nodes).

Regenerated catalog, production solver, no rule: 84.71% / mF1 0.8137 (from 83.47 / 0.8009); held-out +173 conditioned (p < 1e-4), tuning +109, false change 1,437 → 1,150, AKT 100/102, experimental +3, Mitotic G1 −4. Stated plainly: 94% of the held-out gain is DSB Repair and 187 of 199 fixes are DOWN → NORMAL with truth NORMAL. Adopted on exact correctness grounds; the accuracy gain is one pathway's.

Solver rule in this PR: DS_SCC_BREAK_ROLES=<catalyst|assembly|depletion> — the measured record (assembly alone: +214 held-out on the welded catalog; ≈ 0 on the fixed one, i.e. absorbed by the generator fix). Default off, bit-identical (verified on TP53 and DSB networks against the base). Review fixes: closures read the entry state on the pool/flat/influence paths; self-loops are not closures. test/test_scc_break_roles.jl 56.

Re-evaluation of every welded-era A/B on the fixed catalog: eight verdicts stand (elasticity is +29/−85 with or without the welds — it damps the real MDM2–TP53 loop; jacobi, dedup, catalyst-break, own-product not adoptable; depletion floor and hill_sat right; assembly rule redundant); sharing becomes adoptable on structure (accuracy churn-sized); composition edges become clearly harmful (they re-weld: DSB −241).

Records

specs/017-loop-pool/, specs/018-derived-edge-loops/ (spec, plan, tasks, research with P1–P8 pre-registered and scored, two review errata sections each), CI runs the new assertion files, CLAUDE.md test table and knob list updated.

🤖 Generated with Claude Code