Does recycling actually matter at the scale of a real supply risk?

Why this problem, why this architecture

One reusable base class, CohortSurvivalMFA, implements right-censored survival analysis (Kaplan-Meier), Weibull extrapolation, cohort-by-cohort conditional-survival retirement projection, and secondary material accounting, once. A subclass provides exactly one thing, real sourced material content by category, everything else is inherited unchanged. Running two genuinely different technology systems through that same base class, not just one, without modifying it beyond a single additive method, is the actual test of "modular and reusable," not just a claim.

How it's wired

Wind turbines have real individual retirement events to fit a lifetime model from, the Danish register has real decommissioned turbines, verified to reproducedk-wind-mfa's own published numbers exactly (9/9 checks). The global EV fleet does not, it is simply too young: Argonne National Laboratory's own real assessment found only about 3% of global PEV battery capacity had been scrapped as of December 2020. So the EV case sources its Weibull parameters from a real, externally published vehicle survivability table (NHTSA, DOT HS 809 952), the same real source Argonne's own EV assessment uses for exactly this purpose, via one small additive method (fit_weibull_to_curve) rather than a fork of the engine. Material content per vehicle is derived from real cathode chemistry: real chemical formulas (NMC811, LFP) and real atomic masses split a real, reported total cathode mass into individual elements, not an invented split.

A real bug this caught, not a hypothetical one

Building a cross-material comparison chart exposed a genuine mistake: the connector was comparing a multi-decade cumulative recovery total (2025-2050) against crm-trade-network's real one-year disruption-shortfall estimate, two different time scales silently compared as if they were the same thing. For copper the mismatch stayed under 1% and was easy to miss; for graphite, a smaller real global market, cumulative recovery came out to over 1,000% of one year's real shortfall, mathematically consistent with the inputs but not a meaningful claim.

What was actually verified before being trusted

• Every material has a real (low, central, high) content range, not a point estimate, propagated from source through to a real confidence band on the retirement-wave chart.
• Price is given the same treatment where a real market benchmark genuinely matches the traded product (copper, lithium, nickel); where the best benchmark found prices a different grade than the trade code actually blends (cobalt, graphite, a four-element rare-earth basket), that real citation is shown as disclosed context instead of forced into a misleading range.
• crm-trade-network's own 20%-slack cascade parameter has no independent citation, by its own README's admission, so the dashboard cites that gap directly rather than hiding it, and adds one real, independent consistency check instead (Statista's 2023 global copper mine capacity utilization, 77.6%, implying roughly 22.4% real spare capacity).
• Deployment itself was verified, not assumed: an actual docker build plusdocker run caught a real pandas 2.x/3.x incompatibility in a sibling dependency before it ever reached this live demo, and the Streamlit Community Cloud path was checked by genuinely simulating its constraints (a bare container with only this one repo present, no sibling folders, no Docker) before deploying for real.