Equilibrium search over era-5 supply chains

examples/optimize_supply_chain_equilibrium.py loads the transformation matrix directly from era_5_automation/building_levels.csv, trusting the saved input and selecting the hard-coded ALLOWED_CHAIN_GOODS columns. It does not read the adjacent summary.csv. One chain level reproduces a saved recipe targeting £100,000/week of terminal output at base prices. Each trial transforms chain levels into building levels, then solves household wealth, demand, and market prices using the existing equilibrium solver.

Generate inputs with examples/parse.py, examples/production_analysis.py, and examples/supply_chain_compare.py first. The example selects recipes in ALLOWED_CHAIN_GOODS order and includes every configuration used by a selected recipe, including supporting infrastructure outside the terminal graph. It filters unused matrix rows and retains the recipes' production ratios. If an older artifact collection lacks building_levels.csv, rerun uv run python -m examples.supply_chain_compare to generate it. Per-good process reports are no longer used to reconstruct recipes. The example assumes the source recipes use the preset's throughput multiplier of 1; the matrix does not carry throughput metadata. Exported target metadata assumes the preset's £100,000 base-price target; target quantities are derived by dividing that value by each good's base price.

The 20 allowed chains, in vector order, are clothes, furniture, groceries, coal, oil, electricity, glass, services, liquor, tobacco, opium, wine, tea, coffee, transportation, automobiles, luxury_clothes, luxury_furniture, porcelain, and small_arms. Other goods can still be produced by upstream processes or as coproducts within these recipes.

The objective is GDP. The worker cap is one million, the workforce ratio is 0.4. Infrastructure is unconstrained (min_net_infrastructure=None); usage remains a reported diagnostic, and the summary records an empty minimum. The example uses the shared banned building groups, no goods-balance floor, fixed throughput of 1, and no level-dependent economies of scale. The initial vector assigns equal chain levels and scales them to one million workers. Matrix dimensions are derived from the surviving recipes after unused rows are filtered out.

uv run python -m examples.optimize_supply_chain_equilibrium --max-evaluations 100 --max-iterations 10 --log-every 10

Operating losses are funded by default through a proportional tax on cooperative and dependent income, without guaranteeing payroll. Disable it with:

uv run python -m examples.optimize_supply_chain_equilibrium --no-subsidize-operating-losses --max-evaluations 100 --max-iterations 10

The selected policy applies to saved baselines and every restart. --resume uses the current invocation's policy, which defaults to enabled; summary.csv records it in subsidize_operating_losses. Imported calls also default to enabled; use run(..., subsidize_operating_losses=False) to disable it. Unfundable starting or saved mixes still enter the optimizer, whose funding-shortfall penalty guides search toward a fundable recipe mix. An unfunded saved mix cannot become the best solution. Its initial wealth/GDP metrics are empty, and the summary records initial_equilibrium_status=insufficient_income. Only verified, funded equilibria are exported. If no valid result is found within the trial budget, existing exports remain untouched. Other initialization errors still propagate. See tax-funded operating losses.

The default hard trial budget is 20,000 per local search; --log-level WARNING|INFO|DEBUG controls progress verbosity. After each local search, the example reduces loss-making terminal producers' chain levels by --loss-reduction (default 0.5), keeps other levels unchanged, and starts the next search with a fresh trial budget. It does not scale employment back up. The adjustment includes coproduct revenue and uses the existing numerical loss tolerance; upstream profits do not hide terminal-producer losses.

Iteration stops when a search's verified weekly GDP is strictly lower than the preceding search's GDP. Equal GDP continues, with the earliest solution retained on ties. --max-iterations optionally limits the number of completed local searches. Omit it to continue until GDP decreases; an equal-GDP plateau can then continue indefinitely. --loss-reduction 0 disables reductions for every restart.

The example returns the best feasible, freshly verified equilibrium found and writes its solution reports once, after iteration ends. The worse result that triggers stopping is recorded in the summary. A small trial budget can stop during initialization, retaining the baseline with search_success=False. Budget exhaustion does not imply that the equilibrium solver failed or that a global optimum was found. Financial deficits and unaffordable households remain visible diagnostics.

To start the next search from the saved solution and halve loss-making goods' starting chain levels:

uv run python -m examples.optimize_supply_chain_equilibrium --resume --loss-reduction 0.5 --max-evaluations 100 --max-iterations 10

Resume aligns saved levels by good and verifies that the saved recipe matrix matches the current recipes. It re-evaluates the saved solution, then identifies negative operating surplus among each recipe's terminal-good producers, including coproduct revenue. It scales those chains by 1 - loss_reduction and keeps other chain levels unchanged, without scaling employment back up. This changes the starting point; the optimizer can subsequently increase or decrease any allowed chain. The adjusted vector must satisfy scenario constraints. The reduction fraction must be at least 0 and less than 1 and defaults to 0.5. The re-evaluated saved solution remains eligible as the best export. The first new search is compared against this baseline, so a worse first result stops iteration and preserves the saved solution. Use --resume --loss-reduction 0 --max-iterations 1 for a single search starting from the saved vector unchanged. Without --resume, the first search starts from equal chain levels; subsequent searches use the same profitability adjustment as resumed runs. Validation and equilibrium errors propagate before any export is overwritten.

Outputs

The separate collection artifacts/analyses/optimization/supply_chain_equilibrium contains:

• chain_matrix.csv: labeled building-by-chain transformation coefficients.
• chains.csv: invocation starting levels and best chain levels with recipe normalization targets.
• buildings.csv: positive building levels and equilibrium finances.
• market.csv: all goods' buy/sell orders and equilibrium prices.
• pop.csv: population, wealth, income, expenses, and balances by configuration and profession.
• goods_buy_packages.csv: wealth-indexed goods baskets, with wealth retained as a CSV column.
• summary.csv: one row per completed local search, including initial/final metrics, scenario settings, search diagnostics, GDP change, reduction counts, UTC start time, and elapsed seconds. Resumed runs also include iteration zero for the re-evaluated saved baseline, with empty search diagnostics. is_best identifies the exported solution. Every row includes the invocation's completed-search count, total trial evaluations, total elapsed seconds, selected best iteration, optional iteration cap, and stop reason (objective_worsened or max_iterations). Trial totals follow the optimizer's counting convention: initialization counts, but final verification and separate baseline/starting-point evaluations do not.

The committed snapshot was generated with the earlier average_wealth objective. Rerun the example to generate results with the current GDP objective.

objective	annual_gdp	average_wealth	employment	trial_evaluations	search_success	equilibrium_status
gdp	22276396.47676732	27.59560794664505	1000000.0	289	True	converged

Download matrix · Download chain levels · Download building levels · Download market report · Download population report · Download goods baskets · Download summary

Matrix rows use building + "+" + production_method. Columns and chain-level rows use terminal-good keys. Their product reproduces the returned building levels; the building report filters zeros and sorts positive levels. See the guide and column definitions.