California Freight Cleanup → Investigation 8-5

Do better sensors change which portfolio CEC should fund?

$0 VOI at canonical cell • $0.17B in loose-CRF regime • CRF research dominates • 4 strategies × 27 decision cells

Four earlier investigations built increasingly sophisticated sensor strategies — ranking sites by information value, sequencing deployments, running cheaper screening evaluations, fusing satellite and ground data. Each one improved some metric. This investigation asks the harder question: does any of it change which portfolio CEC should fund? The answer is honest and consequential: under the cascade’s actual decision conditions, knowing the PM_2.5 field more precisely does not move the decision. This is not a failure of the sensor work. We built it right; it told us where the real leverage is.

Why this investigation exists

Each prior sensor investigation answered a different technical question. We ranked sites by how much health uncertainty they resolve, optimized deployment sequences, demonstrated cheaper multi-fidelity evaluation, and improved the PM_2.5 field through Bayesian data fusion. Each answer built on the last. None of them answers the question CEC must actually answer: does knowing the PM_2.5 field more precisely change which portfolio we fund?

A strategy that wins on accuracy can produce zero decision value — if the existing uncertainty is already concentrated enough that no information shift crosses the decision boundary. Investigation 8-5 tests this across 108 scenario combinations (three MC budget levels × three decision thresholds × three prior-width settings) for all four sensor strategies.

Methodology

Decision baseline. The Investigation M-1 portfolio frontier defines the decision menu: six named portfolios (A_free_lunch through F_maximum_impact) each with a prior mean net benefit and uncertainty interval derived from the Investigation M-1 MC cascade. The optimal portfolio under the prior is A_free_lunch ($0.68B expected net benefit, σ = 0.264). All alternatives have negative expected net benefit under the prior.

Why A_free_lunch dominates here but D_all_in_4B dominates in Investigation 6-6

The “prior” used in this VOI analysis is the Investigation M-1 MC-cascade single-pollutant baseline (conservative CRF, narrow uncertainty channels). Under that prior, paid portfolios (B–F) have negative expected net benefit because the CRF is small. Investigation 6-6 evaluates the same portfolio set against the Investigation 6-3 multi-pollutant posterior (HR = 1.28, ~4× larger than the Krewski single-pollutant anchor); under that posterior, paid portfolios’ mortality benefits scale up enough to dominate cost, and D_all_in_4B becomes regret-minimizing. Both results are correct against their stated priors. This investigation’s argument is that under the Investigation M-1 prior, sensor expansion cannot move the decision — not that A_free_lunch is the global recommendation.

Strategy-to-channel mapping. Each strategy is mapped onto a set of multiplicative σ-reduction factors for three uncertainty channels driving portfolio net-benefit variance: σ_β (CRF uncertainty), σ_emissions (emission field uncertainty), and σ_surrogate (dispersion model uncertainty). Channel weights combine in quadrature: β = 50%, emissions = 30%, surrogate = 20% (heuristic; a Sobol decomposition would refine these).

Strategy	σ_β factor	σ_emiss. factor	σ_surrog. factor	Portfolio σ multiplier
Investigation 8-2 adaptive placement	1.000	0.136	0.714	0.779
Investigation 8-3 BOCA screening	1.000	0.561	0.144	0.774
Investigation 8-4 sensor fusion	1.000	0.842	0.918	0.939
CRF tightening (meta-analysis)	0.700	1.000	1.000	0.863

VOI computation. VOI(strategy) = E[NB(p* | posterior)] − E[NB(p* | prior)], where each strategy shifts the portfolio-NB variance by the channel factors above. A decision-threshold gate zeros out VOI contributions below the minimum flip-relevance threshold (the smallest net-benefit gap that would change the decision).

27-cell sweep per strategy (108 evaluations total). MC budget ∈ {1,000; 5,000; 20,000}, decision threshold ∈ {$50M; $250M; $1B}, CRF posterior width ∈ {0.5×; 1.0×; 2.0× Investigation 6-3 σ_β}.

The honest negative result

Canonical default cell: all four strategies produce $0 VOI

At the default cell (5,000 MC draws, $250M flip-relevance threshold, 1.0× CRF posterior width), every strategy — adaptive placement, BOCA screening, sensor fusion, and CRF tightening — produces $0.000B headline VOI. Zero.

The Investigation M-1 prior concentrates around A_free_lunch with a $0.68B gap to the next-best alternative B_transport_2B (−$0.58B). Under the canonical CRF posterior width from Investigation 6-3, no realistic σ-reduction from sensor expansion or meta-analysis is large enough to flip the decision above the $250M threshold. The prior dominates. This is the correct answer to the correct question — not a methodology failure.

Strategy	Default-cell VOI ($B)	Max VOI across surface ($B)	Cells dominated (of 27)
Investigation 8-2 adaptive placement	$0.000	$0.148	3
Investigation 8-3 BOCA screening	$0.000	$0.136	3
Investigation 8-4 sensor fusion	$0.000	$0.184	3
CRF tightening	$0.000	$0.071	3

What the negative result means for program investment

VOI for additional sensors as a function of CRF tightness regime; ≈$0 default, $0.17B in loose-CRF regime — Figure: VOI for each sensor strategy across three CRF-tightness regimes. Under default conditions (CRF ×1.0, $250M decision threshold) every strategy earns $0.000B — the prior dominates. Only in the loose-CRF / tight-decision stress regime (CRF ×2.0, $50M threshold) does any strategy produce material VOI: Investigation 8-4 sensor fusion leads at $0.170B, Investigation 8-3 BOCA at $0.131B. This is a robustness check, not a primary recommendation.

This result is only obtainable by running the full decision pipeline end to end. The individual sensor investigations produce positive numbers — EVSI, RMSE improvement, ROI — that look like strong arguments for sensor investment in isolation. Investigation 8-5 adds the decision gate. Under the cascade’s actual uncertainty structure, those improvements do not shift which portfolio CEC should fund.

The programmatic implication is direct: scarce decision-analysis dollars should go to CRF research (Element 6—Investigation 6-3), not sensor expansion. CRF uncertainty (σ_β) carries 50% of portfolio net-benefit variance. Sensor strategies do not touch it. CRF tightening via meta-analysis does — but even that produces only $0.071B max-surface VOI, marginal under current decision conditions. When the CRF posterior widens, CRF research is the lever; sensor expansion is not.

Loose-CRF / tight-decision regime: $0.17B VOI, Investigation 8-4 fusion leads

When the CRF posterior is widened to 2.0× the Investigation 6-3 σ_β and the decision threshold is tightened to $50M, the full VOI surface wakes up: Investigation 8-4 sensor fusion leads with $0.170B VOI; Investigation 8-3 BOCA screening follows at $0.13B. This is the regime where PM_2.5 field precision actually shifts portfolio rankings because CRF uncertainty makes the net-benefit spread narrower. It is a robustness check, not a primary finding.

Regime	Best strategy	VOI ($B)	Runner-up	Gap ($B)
Default cell (canonical)	Tie (all strategies)	$0.000	—	—
Loose-CRF / tight-decision	Investigation 8-4 fusion	$0.170	Investigation 8-3 BOCA	$0.039
Tight-CRF / loose-decision	Tie	$0.000	—	—

A real negative result: the sensor work was right, it just told us something we needed to hear

Every prior Element 8 investigation was built correctly. Investigation 8-1’s VOI map identifies where sensors buy the most health-uncertainty reduction. Investigation 8-2 sequences a deployment programme. Investigation 8-3 demonstrates cheaper multi-fidelity evaluation. Investigation 8-4 improves the PM_2.5 field. All real. All defensible.

What Investigation 8-5 adds is the answer none of those could give alone: does any of this change the decision? Under current cascade conditions, no. Reporting that honestly is what a decision-grade analysis does. This is not a failure of the sensor work — it is the work doing exactly what CEC funded the cascade to do.

Upstream inputs

Investigation	Value consumed
Investigation M-1 (portfolio frontier)	Per-portfolio prior μ, σ NB; deaths mean/SD	`investigations/15_portfolio-frontier/latest/results.json` (sha256 145dbfd826d0)
Investigation 6-3 (hierarchical CRF)	μ_posterior = 0.02439, σ_posterior = 0.004469	`investigations/21_hierarchical-crf-posterior/latest/results.json` (sha256 3104ba850408)
Investigation 8-2 (adaptive placement)	L4 EVSI = $117.8M, EVSI ceiling = $120M	`investigations/27_monitor-adaptive/latest/results.json` (sha256 c92a5b8aface)
Investigation 8-3 (BOCA screening)	BOCA realized EVSI = 2.519, full-sim EVSI = 2.572	`investigations/32_monitor-multifidelity/latest/results.json` (sha256 1b31413e6cd2)
Investigation 8-4 (sensor fusion)	Fusion std = 0.678, model-only std = 0.805	`investigations/38_sensor-fusion-bayesian/latest/results.json` (sha256 f9a7e68e410c)

Caveats

Linear-decision-boundary assumption. Portfolio NB modeled as Gaussian with prior μ/σ from Investigation M-1. Nonlinearity in CRF × emissions × cost is not propagated explicitly.
Channel weights are heuristic. (β = 50%, emissions = 30%, surrogate = 20%). A Sobol decomposition on Investigation M-1 would refine these. The $0 default-cell finding is robust to reasonable reweightings because the A_free_lunch gap ($1.26B) is large.
Sequential RNG. Strategies share one random generator (seed 20260429) and consume draws in order. Absolute VOI levels are not per-strategy IID. Relative ordering within a cell is reliable; the $0 default-cell result is robust across any RNG order since the threshold is not crossed at any budget level.
Single Investigation M-1 decision menu (6 portfolios). A continuous portfolio space (Investigation 6-4 robust frontier) would change absolute VOI magnitudes but not the $0 default-cell finding, since the prior dominance follows from A_free_lunch’s large positive NB relative to all alternatives.
CRF tightening cost ($1M) is order-of-magnitude, from Investigation 6-2 resolution_options. Not a committed budget.
Anti-collapse rule is a methodological commitment: the full 27-cell × 4-strategy surface is always reported. Regime-conditional winners are the deliverable. Suppressing the “tie” cells would be analytically dishonest.

Provenance

Item
run.py	`[internal artifact]`
results.json	`investigations/46_sensor-strategy-cross-validation/latest/results.json` (sha256 e179c5a0ebbb)
analysis.md	`investigations/46_sensor-strategy-cross-validation/latest/analysis.md`
scenario.md	`investigations/46_sensor-strategy-cross-validation/latest/scenario.md`
Howard 1966	Information value theory. IEEE TSSC 2:22–26
Raiffa & Schlaifer 1961	Applied statistical decision theory (EVPI/EVSI structure)
Borenstein et al. 2009	Introduction to Meta-Analysis. Wiley (CRF tightening anchor)
Last run	2026-05-04 (results sha256 e179c5a0ebbb)