California Freight Cleanup → Cross-Cutting Methods

What holds the eight focus areas together?

Five threads of analysis don’t fit cleanly under any one focus area: how the candidate portfolios get built, whether adaptive sequencing pays, what the full Pareto frontier looks like beyond the named portfolios, where the next research dollar should go, and how all the uncertainty propagates from one investigation to the next. Three remain current. The research-priority synthesis and uncertainty-propagation attempt are withdrawn and retained only as audit history.

→ Decision Dashboard — compare portfolios across CRF anchors and budget scales.

The five threads

Investigation M-1 — Building the candidate portfolios

The starting point. We assemble roughly 120 combinations of transport, building, and wildfire interventions and name six representative portfolios from $0 B (the “free lunch”) to $13.9 B (max impact). Every downstream investigation operates on these. Headline finding: at the $2 B marginal, transport electrification avoids 66.8 deaths per year — 5.8× the return of wildfire fuel management at the same dollar.

Investigation M-2 — Does adaptive sequencing pay?

If you can re-plan year by year as health data comes in, instead of committing to a 10-year budget upfront, do you save more lives? Yes, by about +27.1% (1,692 vs 1,332 deaths over a $4 B, ten-year horizon) — but with an honest caveat. The full belief-conditioned adaptive policy collapsed to a single dominant action in our setup. The gain we’re seeing is from rolling-horizon re-optimization, not true belief adaptivity. We say so in the writeup.

Investigation M-3 — The full Pareto frontier

The named portfolios in Investigation M-1 collapse health, equity, and cost into a single net-benefit score. That hides an implicit exchange rate between disadvantaged-community (DAC) deaths and non-DAC deaths. NSGA-II refuses that collapse: we let it search a 5-dimensional design space and find a 100-point Pareto frontier across all three objectives. Result: 48 of those points strictly beat E_smart (the equity-targeted hand-crafted portfolio) on every dimension. The best one saves 580 more lives at $0.23 B less. The lever pulling the equity corner of the frontier is indoor air quality — a control absent from the original menu.

Investigation M-4 — Historical research-priority synthesis

This synthesis combined Investigation 3-10’s joint sensitivity indices with the cost of resolving each input. Because Investigation 3-10 now has stale upstream provenance and is withdrawn, M-4’s ranking and value-of-information estimates are withdrawn too. The old page remains visible as a historical snapshot; a clean parent rerun and fresh synthesis are required before this thread can guide a research budget.

Investigation M-5 — Withdrawn uncertainty-propagation attempt

This investigation tried to trace node-specific uncertainty through a shortened analytical chain. A later code review found that independent random streams could create an apparent shift even when a perturbation had no effect, and that some advertised adapters did not reach the terminal calculation. Its numerical results and robustness conclusion are withdrawn. The historical page remains available for audit while the method awaits a corrected implementation, clean rerun, and human review.

Methodological assumptions threaded through the cascade

CRF shape: log-linear (no-threshold) form is assumed throughout the cascade per EPA 2023 PM_2.5 TSD. Supralinear (Burnett GEMM 2018) and threshold (Goodkind 2019) shapes are not evaluated; only β-parameter uncertainty is propagated, not shape uncertainty.
CRF national-to-California transfer: Di and Krewski cohorts are US-wide; their hazard ratios are applied to California under the assumption that the dose-response is identical at the state level. California-specific re-estimation lives in Investigation 6-3 (which produces a higher β, attributed to region-mean exposure attenuation; see Investigation 6-2 for the reconciliation).
Cross-investigation uncertainty: no current joint sampler supports a cascade-wide dependence claim. The former independent-channel implementation is withdrawn, so current results keep their uncertainty statements within each investigation rather than presenting a combined tail estimate.
EVSI factorizations: Investigation 6-2 / Investigation 6-5 use the product form p_success × shrinkage × residual_EVSI, which is an exact identity only under Normal-conjugate updates with linear payoffs. The log-linear CRF introduces a mild concavity; treat ROI multipliers as upper bounds on what a properly decision-theoretic computation would yield.

Headline findings

Transport dominates the $2B sector choice by 5.8×: At the $2B marginal, transport electrification (T2 over T1) avoids 66.8 deaths/yr versus 45.4 for building and 11.5 for wildfire treatment (Investigation M-1).
48 NSGA-II portfolios strictly dominate E_smart on all three objectives simultaneously; best dominator gains 580 deaths at $0.23B less cost. The indoor AQ lever, absent from Investigation 6-4’s hand-crafted menu, drives the high-equity corner of the frontier (Investigation M-3).
Sequential adaptivity is worth +27.1% in cumulative deaths avoided over a static one-shot allocation, but PBVI policy collapses to a single dominant action (alloc_kr) at every belief state—the gain is best-static-allocation-under-value-iteration, not belief-conditioned adaptivity (Investigation M-2).
No current Sobol-weighted research-priority ranking is claimed. Investigation M-4 is downstream of the withdrawn Investigation 3-10 result and remains historical until both analyses are rerun cleanly.
No current M-5 propagation result is claimed. The historical implementation did not isolate perturbation effects reliably and did not connect every advertised adapter to the reported terminal outcome. Its former robustness statement is withdrawn.

Investigations

Inv	Slug	Tier	One-line	Deep-dive
15	portfolio-frontier	Tier 2	Efficient frontier + 6 named portfolios; $2B head-to-head winner = transport (5.8× wildfire)	Deep-dive
22	portfolio-pomdp	Tier 2	5-level sequential policy ladder; PBVI L4 +23.9% over one-shot; half-normal cost-overrun baked in	Deep-dive
29	nsga-pareto-equity	Tier 2	NSGA-II 3-objective Pareto (health × equity × cost); 48 frontier points dominate E_smart	Deep-dive
59	sobol-research-priority	Historical	Withdrawn research-priority synthesis; downstream of Investigation 3-10’s stale provenance	Audit snapshot
61	cross-investigation-mc-propagation	Historical	Withdrawn uncertainty-propagation attempt; retained to document the failed method and the requirements for a valid rerun	Audit snapshot

Cross-element notes

Investigation M-1 feeds Element 6 (Ratepayer Burden). The six named portfolios defined here are the candidate set consumed by Investigation 6-4 (robust optimization) and Investigation 6-6 (CRF-conditional decision), both of which are the analytical backbone of the Element 6 ratepayer-burden page. If the Investigation M-1 frontier composition shifts, Investigation 6-4 and Investigation 6-6 must re-run.

Investigation M-2 cost-overrun model feeds Element 6. The half-normal cost-overrun distribution (mean 1.16×, P95 1.392×) calibrated in Investigation M-2 is also consumed by Investigation 6-6 (Phase 3 regret-surface computation) and the now-withdrawn Investigation 3-10 historical joint test. This investigation quantifies program-execution risk — the number any ratepayer-burden calculation needs before it is defensible.

Investigation M-3 NSGA candidates feed Element 6 via Investigation 6-4 and Investigation 6-9. The Q_nsga_1 and Q_nsga_2 portfolios identified by NSGA-II are imported into Investigation M-2 as external comparators, and the Q_nsga_2 portfolio ($1.78B indoor-heavy) was evaluated by Investigation 6-9 (ISRM validation), which returned VERDICT AMBIGUOUS under the Di CRF. That caveat is carried forward in the Element 6 narrative.

Investigation M-4 no longer refines the current Element 6 roadmap. Its Sobol-weighted synthesis depends on the withdrawn Investigation 3-10 parent and is therefore historical. The current component analyses can still motivate research questions, but they do not support M-4’s joint ranking.

Investigation M-5 does not currently support a stress-test claim. Its historical implementation confounded perturbation effects with Monte Carlo noise and allowed broken adapter links to fall back silently. The retained page documents why the result was withdrawn and what must change before the question can be answered defensibly.