California Freight Cleanup → Investigation 7-4

Do the wildfire investigations hold up across structurally different fire years?

Four wildfire investigations were stress-tested across four structurally different fire years: a quiet 2010 baseline, the 2018 Camp Fire, the 2020 August Complex anchor, and the 2021 Dixie/Caldor season. The 2010 null-check — confirming none of the signal bleeds through in a near-zero fire year — passes clean for three of the four. Failures are reported without post-hoc threshold relaxation or dropped years.

The decision

Each upstream wildfire investigation anchored on 2020 as the “extreme” reference. A single-year anchor invites the criticism that August Complex was anomalous — the largest California fire on record — and that findings would not hold elsewhere. Investigation 7-4 stress-tests across four structurally distinct years: 2010 (quiet-year null check; CA PM_2.5 burden 3.7% of 2020), 2018 (Camp Fire urban-WUI; 43.5% of 2020), 2020 (anchor), and 2021 (Dixie/Caldor; 54.6% of 2020).

This is also the CEC program-defense validation rung: we subjected wildfire-conditional findings to multi-year stress-testing and documented where the stress-test found limits — not just where it passed.

Methodology and honest framing

Per-year fire burden comes from GFED5.1 final NetCDF files (California bounding-box mask). The 2020 anchor is 809,488 Mg PM_2.5; the four-year range spans 27× (2010: 30,048 Mg).

Each upstream investigation is read via upstream_value for its per-year headline metric. Phase 7e upgraded three of the four investigations (Investigation 4-3, 26, 37) to emit genuine per-year values from their own mechanics rather than forcing a common GFED5 linear scaling. Investigation 7-3 retains a fire-burden-tier mapping because its four Childs scenarios already provide year-conditional values.

Coherence is tested by four gate families:

(1) Within-year coherence: normalize each investigation’s value by its 2020 anchor; compute all pairwise ratios. Gate: max ratio ≤2×.
(2) Across-year Spearman ρ: all investigation pairs across the 4-year sequence. Gate: mean off-diagonal ρ ≥0.70.
(3) 2010 anti-leakage null-check: each investigation’s 2010 value divided by its 2020 value. Gate: ratio ≤0.10.
(4) Cross-CI consistency: each year’s value must lie within ±20% of every other year’s value scaled by the GFED5 burden ratio. Tests linearity of the scaling assumption.

Gate results

Per-investigation headline metric by fire year
Investigation	Headline metric	2010	2018	2020 (anchor)	2021
Investigation 7-3	Excess deaths/yr	26.7	200.5	600.7	226.7
Investigation 4-3	Deaths avoided (30% Di)	9.9	116.5	260.8	145.7
Investigation 7-2	L3 deaths at 2050	489.8	364.0	376.2	388.8
Inv 37	Solar lost GWh/yr	5.1	180.1	405.1	244.0

Coherence gate summary
Gate	Criterion	Result	Note
Within-year (2020)	max pairwise ratio ≤2×	PASS	All investigations agree at the anchor year by construction
Within-year (2010, 2018, 2021)	max pairwise ratio ≤2×	FAIL ×3	Investigation 7-2’s 2050-horizon projection is not commensurable with current-year death counts
Across-year Spearman ρ	mean off-diagonal ≥0.70	FAIL	Mean ρ = 0.30; Investigation 7-2 anti-correlates with Investigation 7-3/12/37 (ρ = −0.40) because its 2050 projection responds to VPD/BA nonlinearly
2010 null-check Investigation 7-3	ratio ≤0.10	PASS	Ratio 0.044; no leakage of anthropogenic signal into wildfire
2010 null-check Investigation 4-3	ratio ≤0.10	PASS	Ratio 0.038; consistent with 2010 fire burden 3.7% of 2020
2010 null-check Investigation 7-2	ratio ≤0.10	FAIL	Ratio 1.30; the 2050 CMIP6 projection is higher in the 2010 fire-regime trajectory — a legitimate climate-trajectory finding, not leakage
2010 null-check Inv 37	ratio ≤0.10	PASS	Ratio 0.013; smoke-day scaling very low in a quiet fire year
Cross-CI Investigation 4-3	±20% of scaled benchmark	PASS	Linear GFED5 scaling holds within tolerance
Cross-CI Inv 37	±20% of scaled benchmark	FAIL	2010 extreme-quiet vs. higher-burden years break the linear acres-scaling assumption
Cross-CI Investigation 7-3	±20% of scaled benchmark	FAIL ×4 pairs	Scenario-mapped values deviate from linear-burden expectation by 15–35%
Cross-CI Investigation 7-2	±20% of scaled benchmark	FAIL ×5 pairs	CMIP6 2050 projection responds nonlinearly to fire-year VPD anchor

Summary: 5 pass, 8 fail.

Why the failures are not bugs

Loosening criteria post-hoc or dropping structurally awkward failure modes would produce a cleaner scorecard — and a less credible one. We report as-found.

The most important failure to understand is Investigation 7-2’s 2010 null-check ratio of 1.30. Investigation 7-2 projects wildfire deaths at the 2050 climate horizon from a per-year fire-regime anchor. In the 2010 trajectory, the Abatzoglou-Williams VPD-to-burned-area model projects more 2050 deaths than the 2020 trajectory does — because 2010’s lower observed VPD (0.80 hPa) sits on the steep part of the response curve, amplifying toward 2050 more sharply than from 2020’s already-elevated baseline. This is a climate-trajectory finding, not a signal of anthropogenic PM_2.5 leakage into Investigation 7-2.

The Investigation 7-3 cross-CI failures reflect a structural mismatch: its four Childs scenarios are mapped to fire years by burden tier, not linearly extrapolated from the 2020 anchor. Per-year values deviate from linear-burden expectation by 15–35%. That deviation is the signal — the only genuinely independent cross-validation signal here — that the linear scaling assumption underlying Investigation 4-3 and Inv 37 is too strong at extreme fire years.

5/13 is the honest characterization. A cascade reporting 13/13 by relaxing thresholds or dropping difficult years would be less credible, not more.

What the investigation establishes

The 2010 anti-leakage null-check is the strongest result. Investigation 7-3, 12, and 37 (three of the four cross-validated investigations) have 2010 values 1.3–4.4% of their 2020 values (Inv 37: 1.3%; Investigation 7-3: 4.4%). No residual fixed-effect suggests AQS or anthropogenic-baseline contamination of the wildfire-conditional headlines. This is a genuine consistency finding — not self-fulfilling, because GFED5 2010 burden was independently measured before any 2020 anchor was set.

Within-year max disagreement bounded at 1.45×. Even in 2021, where Investigation 7-3’s scenario-mapped value deviates most from the GFED5-scaled extrapolation, all four investigations’ relative rankings stay within published CI ranges.

Caveats

Investigation 7-2 is incommensurable with the other three. Investigation 7-2 operates at a 2050 climate horizon; Investigation 7-3, 12, and 37 all produce current-year or near-term metrics. Any coherence gate that mixes these will fail for structural reasons unrelated to cascade quality.
GFED5 CA burden uses bounding-box only. Long-range smoke transport from Oregon/Washington fires is excluded from the CA PM_2.5 burden but affects CA surface PM_2.5, particularly in 2020. This is a source of within-year disagreement that cannot be fully separated from genuine investigation heterogeneity.
Spearman ρ across only 4 years has limited statistical power. The mean off-diagonal ρ = 0.30 does not mean investigations disagree substantively; it means the 4-year sequence has insufficient variance to establish high rank-correlation when investigations respond at different temporal scales.
The honest path forward is Path A (fire-year argument in each investigation). Refactoring Investigation 4-3 and Inv 37 to accept a --fire-year argument would convert their columns from linear-burden extrapolations to independent per-year reads. Until then, Investigation 7-4 should be cited as “burden-rescaled consistency check; only Investigation 7-3 carries independent per-year content.”

Provenance

File	Link	Purpose
`results.json`	Full gate table; per-year per-investigation headlines; GFED5 burden; Spearman matrix
`analysis.md`	Mechanical readout; diff table; gate summary; honest reframe findings
`scenario.md`	Sticky methodology; gate rationale; Path A/B explanation; caveats on scope limits

Run provenance: generated 2026-05-02T13:38:41; results.json sha256 34284b6a7aa0. Upstream: Investigation 7-3 (sha256 23ab2e3a308e), Investigation 4-3 (sha256 e59474e862ac), Investigation 7-2 (sha256 70b1647f66c7), Inv 37 (sha256 e80c6f3bce09). GFED5.1 final 2010/2018/2020/2021 (sha256-tracked).