feat(napkin-math): shared-pool legitimacy check for aggregate surplus tests

[neoneye]

Summary

ChatGPT reviewed the v31 cross-border rail ticketing run (the cross-plan generalisation test from PR #706) and flagged one real modelling discipline issue:

The combined_program_viability_surplus_eur may be too aggressive or conceptually overloaded. It mixes liquidity/clearing solvency with revenue loss from enforcement delay. Those may not draw on the same reserve unless the report explicitly says the emergency float reserve also backstops regulatory/revenue shortfalls.

The rail digest explicitly scopes the €300M emergency reserve to the clearing mechanism — those two pressures draw on different pools. Additive netting overstated headroom. Conceptually correct form is min() over the individual surplus calculations.

What changed

Both extractor prompts (extract-parameters-from-full, extract-parameters-from-digest) gain a Shared-pool legitimacy check for combined surplus tests rule:

• When an aggregate test subtracts multiple pressures from one reserve, verify those pressures actually draw from the same pool — same named buffer, same line item, same envelope.
• If yes (single pool absorbs every pressure): additive form is correct.
  • combined_viability_surplus = pool_reserve - pressure_a - pressure_b - pressure_c
• If no (separate pools): use min() over individual surpluses.
  • combined_viability_surplus = min(surplus_a, surplus_b, surplus_c)
• Concrete signs of each pattern listed. Aggregate name should match the form: prefer weakest_gate_surplus / worst_case_pool_surplus when using min().

Cross-validated against both reference plans

Nuuk Clay Workshop: digest explicitly says the same 15% DKK contingency absorbs labor-law, rental, and kiln-failure shocks. Additive form (contingency - labor_shock - rental_shortfall) is correct under the new rule. No model change needed.

Cross-border rail ticketing: v32 replaces additive combined_program_viability_surplus_eur with:

weakest_financial_gate_surplus_eur = min(
  clearing_capacity_surplus_eur,
  regulatory_risk_buffer_surplus_eur,
  royalty_cost_coverage_surplus_eur,
)

Numerical effect of the framing change

Metric	v31 (additive)	v32 (min)
Aggregate pass probability	47.33% (FRAGILE)	26.15% (FRAGILE, near DOOM)
Base scenario value	+87.5M EUR	0 EUR
Low scenario value	+400M EUR	−500k EUR
High scenario value	−800M EUR	−600M EUR

The other five individually-scoped gates are unchanged. The additive form was netting deficits in revenue economics against surplus in clearing liquidity — different stakeholders, different pools. The min() form correctly says "all three gates must independently hold".

Test plan

• Smoke 8/8 (tests/run_smoke.py)
• Unittest 45/45 (tests/test_run_monte_carlo.py)
• v32 rail regenerated with min() framing; runner emits zero warnings; insights.md reads cleanly with the new headline
• Nuuk's additive form re-validated against the digest text — no change needed
• Reviewer: confirm the prompt rule's min() direction is the right default when the source is ambiguous (current text leans toward min() unless the source explicitly names a shared pool — fail safe).

🤖 Generated with Claude Code