PEP-1-1 Parity Coverage Matrix¶

Coverage for the pep2 dataset — the PEP-1-1 v2.1 CGE model ported to Pyomo (equilibria.templates.pep_pyomo). Every row compares the Pyomo solve against a GAMS reference solved by the same engine (IPOPT vs IPOPT, or PATH vs PATH), so the solver’s equality tolerance cancels and the cell-by-cell match reflects model fidelity, not solver noise.

match is the measured snapshot; the named pytest gate re-measures and enforces itlocal every row needs PATH/IPOPT + GAMS refs — no solver-free CI gate (the PEP reference is a solved GDX, not a .nl dump)

NLP vs NLP¶

The Pyomo model solved as an NLP (IPOPT on the raw model, nlp_scaling_method=none, faithful to GAMS’s raw solve) against the GAMS CNS reference Results.gdx. The benchmark BASE reproduces the SAM, so the seeded point is the calibration answer. Run: phase1_nlp.py --model pep --dataset pep2 --period BASE.

Scenario · form	Cells	Match	Gate	Reference
basenlp	317	100%local	phase1_nlp.py --form nlp / test_faithful_at_benchmark	Results.gdx (GAMS CNS, val*)

MCP vs MCP¶

The Pyomo model solved as a complementarity problem via PATH against the GAMS-native MCP reference (PEP-1-1_v2_1_mcp_solve.gms: MODEL /ALL/ + SOLVE USING MCP, so GAMS infers the equation↔variable pairing). The sim1 row is the reference counterfactual — a 25% export-tax cut (ttix.fx=ttixO*0.75) — applied faithfully in Python by scaling the ttixO benchmark before build. Run: phase1_nlp.py --model pep --dataset pep2 --form mcp.

Scenario · form	Cells	Match	Gate	Reference
basemcp	367	100%local	test_mcp_matches_gams_native_mcp	Results_mcp.gdx (GAMS /ALL/ MCP)
sim1mcp	314	100% (GDP_BP 46707→46748.2)local	test_mcp_sim1_shock_matches_gams	Results_mcp_sim1.gdx (GAMS MCP, ttix·0.75)

NLP↔MCP mirror¶

The two Pyomo forms, solved from the same feasible benchmark seed, land on the identical point (LEON, the form-defining Walras slack, is excluded). The one historical gap — PD['othind'], filled from its *O benchmark (1.132) rather than a blind 1.0 — closed the mirror to a clean 100%.

Scenario · form	Cells	Match	Gate	Reference
baseboth	466	100%local	test_nlp_mcp_mirror	self (NLP vs MCP, LEON excl)

objdef variant¶

The objdef variant adds a dummy objective (OBJDEF: OBJ==0, minimize OBJ) — the SOLVE NLP MINIMIZING OBJ lineage. A constant objective cannot move the equilibrium, so objdef-NLP lands on the exact base-NLP point; in the MCP form OBJ is not declared, keeping the system square.

Scenario · form	Cells	Match	Gate	Reference
basenlp	467	100% (== base-NLP)local	test_objdef_variant_equals_base_nlp	self (objdef vs base)
basemcp	358	square, code=1local	test_objdef_mcp_is_square_and_solves	self (objdef+MCP squareness)