California Freight Cleanup → Investigation 6-8
Does the same pipeline give the right answer at a smaller budget?
Q3-B: P(optimal) = 1.000, $3.02M/death, E[regret] = $0.000
The same three-rung pipeline as Investigation 6-4 + Investigation 6-6, applied at the CARB AB 617 cycle scale. At $300M, one intervention slot is affordable after claiming free-lunch — B4 cooking-first wins it. At $4B, multiple sectors become simultaneously affordable and breadth beats concentration. This is not a methodology artifact. The right answer genuinely changes with scale.
The decision
Phase 1 demonstrated the pipeline at the $4B CEC GFO envelope. The central question for ADM applicability: does the methodology generalize to other scales, or is it load-bearing on a particular budget? A pipeline that only works at one decision scale is a custom analysis for one program — not a reusable methodology.
The $300M scale is chosen because CARB’s AB 617 Community Air Protection Program operates at $200–400M cycle scale — the same range as CEC IIP solicitations and California’s Cap-and-Invest Q3 funding rounds. A regulator at AB 617 cycle scale would benefit from the same analytical rigor as Phase 1, applied at the appropriate envelope.
The expected finding — that the pipeline produces defensible recommendations that differ from Phase 1 in ways reflecting genuine scale-dependence — is itself the result. Identical recommendations at both scales would signal budget-insensitivity: a concerning artifact, not a validation.
Methodology
Same three-rung framework as Phase 1’s Investigation 6-4 + Investigation 6-6, applied to five Q3-scale candidates:
| Portfolio | Composition | Cost | Deaths/yr |
|---|---|---|---|
| Q3-A | T1+B1+B3+DTE retire (all $0 free-lunch) | $0 | 75.8 |
| Q3-B | Free baseline + 0.6×B4 cooking-first | $300M | 99.5 |
| Q3-C | Free baseline + 0.3×T4 equity-targeted transport | $300M | 90.3 |
| Q3-D | Free baseline + 0.15B B4 + 0.15B wildfire-5pct | $300M | 88.2 |
| Q3-E | 0.6×B4 alone, NO free baseline (dominated reference) | $300M | 23.6 |
Linear scaling assumption: 60% of the full B4 program ($0.3B of $0.5B) yields 60% of full B4 deaths avoided. This is approximate — a scale-aware analysis requires Investigation 1-2’s per-household marginal CE curve (see Caveats). Q3-E is a deliberately dominated reference, included to verify the framework correctly penalizes omitting the zero-cost baseline.
L0 — Cost-effectiveness ($M per death avoided)
Q3-A is free — dominant by definition. Among paid candidates: Q3-B achieves $3.02M/death, Q3-C $3.32M/death, Q3-D $3.40M/death. Q3-E (no free baseline) costs $12.69M/death — 4× worse than Q3-B. Omitting the free-lunch portfolio is a dominant mistake even within a single sector.
L1 — Expected-value MC (5,000 draws)
CRF CV = 0.16, cost CV = 0.25, effectiveness CV = 0.30; NB = deaths × VSL ($11.6M) − cost. Q3-A wins on raw EV (NB = +$0.930B) because zero cost makes the entire health benefit pure monetized value. Q3-B is second (+$0.911B, P(NB > 0) = 0.997). Q3-E is nearly zero (NB = −$0.020B, P(NB > 0) = 0.403) — dominated status confirmed.
Posterior-integrated regret (mirrors Investigation 6-6)
4,096 Sobol quasi-MC draws over Investigation 6-3’s hierarchical β posterior; 10-year NPV at 3% discount; cost amortized over 10 years. Q3-B achieves P(optimal) = 1.000, E[regret] = $0.000. Q3-A carries E[regret] = $2.09B — the opportunity cost of not funding B4 when the CRF posterior implies B4 is cost-effective at $3.02M/death. Q3-C carries $0.90B regret.
Findings
| Portfolio | NPV at posterior mean ($B) | E[regret] ($B) | P(optimal) | $M/death |
|---|---|---|---|---|
| Q3-A (free only) | +7.52 | 2.09 | 0.000 | free |
| Q3-B (B4 + free) | +9.61 | 0.00 | 1.000 | $3.02M |
| Q3-C (T4 + free) | +8.70 | 0.90 | 0.000 | $3.32M |
| Q3-D (B4+wildfire + free) | +8.49 | 1.12 | 0.000 | $3.40M |
| Q3-E (B4 alone, dominated) | +2.09 | 7.52 | 0.000 | $12.69M |
Scale-dependent composition finding
$300M and $4B call for different strategies — that’s a genuine insight, not a methodology artifact
At $4B (Phase 1 CEC GFO envelope): D_all_in_4B wins — broad, multi-sector, spanning transport, building, wildfire, and DTE retirement (Investigation 6-6, P(optimal) = 1.000).
At $300M (CARB AB 617 cycle scale): Q3-B wins — concentrated on the single highest-marginal-value intervention (B4 cooking-first) stacked on zero-cost free-lunch.
The mechanism is straightforward. At $300M, only one paid intervention slot is affordable after claiming free-lunch; B4 cooking-first fills it at $3.02M/death. At $4B, multiple sectors become simultaneously affordable, and portfolio breadth generates larger aggregate benefit than any single-sector concentration.
This is not a methodology artifact — it is genuine scale-dependence in the underlying decision structure. Each answer built on the last; the recommendations differ because the right answer at each scale is genuinely different.
Cross-scale comparison
| Dimension | Phase 1 ($4B, Investigation 6-4 + 44) | Q3 ($300M, Investigation 6-8) |
|---|---|---|
| L0 best paid | Q_nsga_5 at $2.44M/death | Q3-B at $3.02M/death |
| Regret-min | D_all_in_4B (P(opt) = 1.000) | Q3-B (P(opt) = 1.000) |
| Winning composition | Broad: transport + building + wildfire + DTE | Concentrated: free-lunch + B4 cooking-first |
| Free-lunch included | Yes (within D_all_in_4B) | Yes (explicit in Q3-B) |
| DAC share (winner) | 25.0% (D) | 27.0% (Q3-B) |
Caveats
- Linear scaling at $300M. Funding 60% of B4 ($0.3B of $0.5B) is assumed to yield 60% of B4 deaths avoided. Real CARB IIP programs have minimum scales below which marginal effectiveness drops (community engagement, infrastructure deployment). A fully rigorous analysis would use Investigation 1-2’s per-household marginal CE curve at the $300M scale.
- Q3-E is a dominated reference, not a real candidate. Included to verify the framework discriminates correctly — Q3-E’s $12.69M/death and 40% P(NB > 0) confirm that omitting free-lunch is strictly worse.
- DAC shares are Phase 1 estimates. Carried over from Investigation 5-1 (geographic decomposition at L1). A proper Q3-scale DAC analysis would re-run Investigation 5-1 at the $300M envelope; deferred.
- 10-year NPV horizon vs. 3-year AB 617 cycle. The posterior-integrated regret uses a 10-year NPV at 3% discount. Re-running with a 3-year amortization (the actual AB 617 cycle) would likely shift regret-min from Q3-B to Q3-A, because the higher annual cost-equivalent at 3 years reduces Q3-B’s NPV advantage.
- Comparison to Phase 1 is a meta-finding, not a normative preference. Both case studies operate at different decision contexts (AB 617 cycle vs. CEC GFO). The cross-scale comparison demonstrates methodology generalizability, not that one program is preferred over the other.
- Stale upstream flags at last run. Investigation 6-4 and Investigation M-1 sha256 hashes had changed since Investigation 6-8 last ran. Key outputs (Q3-B winner, $3.02M/death, P(opt) = 1.000) are unchanged in the diff table.
Provenance
| Field | Value |
|---|---|
| Investigation | 52_q3-300m-case-study |
| Tier | Tier 1 |
| Run timestamp | 2026-05-04T07:48:06 |
results. This page is generated from
investigations/52_q3-300m-case-study/latest/results.json
(sha256 63533b8016b4).
|