[adamjohnwright]

bind_output_slots binds an output set to an input set only when reference signatures match exactly, so a reaction like "MDM2 forms homo- or heterodimers" (R-HSA-6804741) was emitted as the full input × output cross product. Two of its four copies are unbalanced: p-MDM2 alone makes MDM2:MDM4 (creating MDM4), and MDM4 + p-MDM2 makes the MDM2 dimer (losing MDM4).

balance_choices (in src/variant_emission.py) scores each copy by the proteins its outputs gain plus the proteins its inputs lose. It drops a copy only when a lower-scoring copy exists with the same outputs AND another exists with the same input choice. So an input with no balanced partner keeps its best copy.

Measured (deltasignal specs/101, registered gate as specs/077), arm bal vs canonical defaults:

• Curator held-out: +30 (34 fixed / 4 broken, p < 1e-4). Mostly IL-4/13: p-STAT6 no longer makes the STAT3 dimer.
• Curator tuning: −12.
• Experimental: −4 (2 / 6). No RAF case moves.
• Worst pathway: ERBB2 −7. EGFR and ERBB3 knockouts no longer feed every ERBB2 heterodimer.
• The MDM4 knockout case inverts: the freed MDM2 homodimer raises the ligase pool.
• Gate passes. Adopted by Adam 2026-10-24.

Structure:

• 38 of 92 pathway tables change.
• The flag-off build is identical to canonical.

Changes:

• Flag LNG_BALANCE_CHOICES, default 1.
• PIPELINE_SETTINGS.md row.
• Three tests in tests/test_variant_emission.py.

Checks: full suite passes except the known stale test_main_edges_proportional_to_best_matches[...983169], which also fails on main. mypy is clean.

🤖 Generated with Claude Code