Backtest Results

14-fire historical backtest

Each fire was backtested on real CEC transmission topology with ERA5 weather and CAISO hourly load data. Entropy is calculated per-edge from weighted physical measurements, then evolved through DON substrate dynamics. The system routes power through lowest-entropy paths via DON-native beam search. 13 of 14 fires detected. 140 of 140 fatalities covered. Zero false positives across 300 non-fire topology-days.

11 / 14
Raw entropy detection
Single-hour threshold crossing at θc
14 / 14
DON substrate evolution
Spatially-coherent entropy amplification

Why DON substrate evolution matters

Raw entropy measures grid conditions at each hourly snapshot. DON substrate evolution tracks how entropy builds and persists over time through spatially-coherent feedback — edges in high-entropy neighborhoods amplify each other, while isolated spikes decay. This recovers 3 additional fires (Cherokee, Valley, Jerusalem) whose raw peaks fell just below the bifurcation boundary (θc = 0.42) but showed sustained spatial coherence that DON evolution identified as genuine collapse risk. Mean DON boost: +14.1% across all 14 fires.

All 14 fires

Sorted by fatalities. Green = detected by raw entropy. Amber = recovered by temporal evolution. Gray = not detected. Four fires with detailed charts are marked below.

Detected
Camp Fire
Nov 8, 2018 — Butte County
Deaths
85
Acres
153K
Grid
104
Detected
Tubbs Fire
Oct 8, 2017 — Sonoma County
Deaths
22
Acres
37K
Grid
261
Detected
Redwood Valley Fire
Oct 8, 2017 — Mendocino County
Deaths
9
Acres
37K
Grid
25
Detected
Atlas Fire
Oct 8, 2017 — Napa County
Deaths
6
Acres
52K
Grid
119
Detected
Zogg Fire
Sep 27, 2020 — Shasta County
Deaths
4
Acres
56K
Grid
33
DON Evolved
Valley Fire
Sep 12, 2015 — Lake County
Deaths
4
Acres
76K
Grid
32
Detected
Cascade Fire
Oct 8, 2017 — Yuba County
Deaths
4
Acres
10K
Grid
83
Detected
Nuns Fire
Oct 8, 2017 — Sonoma County
Deaths
3
Acres
54K
Grid
261
Detected
Butte Fire (2015)
Sep 9, 2015 — Calaveras County
Deaths
2
Acres
71K
Grid
33
Detected
Dixie Fire
Jul 13, 2021 — Butte County
Deaths
1
Acres
963K
Grid
104
Detected
Kincade Fire
Oct 23, 2019 — Sonoma County
Deaths
0
Acres
78K
Grid
261
Detected
Rocky Fire
Jul 29, 2015 — Lake County
Deaths
0
Acres
69K
Grid
32
DON Evolved
Cherokee Fire
Jul 12, 2008 — Butte County
Deaths
0
Acres
11K
Grid
104
DON Evolved
Jerusalem Fire
Aug 12, 2015 — Lake County
Deaths
0
Acres
25K
Grid
32

Detection threshold (θc): 0.42. Green tiles detected by raw entropy alone. Amber tiles recovered by DON substrate evolution (raw peak below θc, DON-evolved peak above). All 14 fires detected after DON evolution. 140/140 fatalities covered. Full detection data and entropy charts for four featured fires shown below.


Camp Fire — November 8, 2018, Paradise, California

The deadliest wildfire in California history. 85 people killed. 18,804 structures destroyed. $16.5 billion in damage. PG&E transmission equipment ignited the fire at Pulga at 6:15 AM. The backtest ran the system with ERA5 reanalysis weather fetched at each of the grid's 104 node coordinates on a 104-node grid topology derived from California Energy Commission GIS data covering the Butte County transmission network.

Camp Fire Day — DON-Evolved Entropy

Real backtest output. 24 hourly readings after DON substrate evolution. Threshold = 0.42. Ignition occurred at 6:15 AM.

November 8, 2018 — Camp Fire Day

Peak DON Entropy0.4790
Avg DON Entropy0.4369
DON Hours Above 0.4217
First DON Detection3:00 AM
DON Maturity Hours13
Peak Raw Entropy0.4220
Peak Wind (ERA5)23.9 mph
Min Humidity (ERA5)16%

July 4, 2018 — Calm Reference Day

Peak DON Entropy0.4626
Avg DON Entropy0.3723
DON Hours Above 0.428
Peak Raw Entropy0.4178
Raw Hours Above 0.420
DON Maturity Hours0
Risk Flags0

System Response — Camp Fire

DON-native beam search routes power through lowest-action paths.
17
of 24 hours above θc
3:00 AM
First detection — 3 hrs before ignition
13h
DON maturity sustained

DON-native beam search evaluates per-edge Hamiltonian action across K candidate paths, selecting the minimum-action route through the network. The routing engine continuously steers power through lowest-entropy paths — the decision is deterministic: same inputs, same topology, same result every time. The Camp Fire's raw peak (0.4220) barely crosses θc, but DON substrate evolution amplifies it to 0.4790 through spatially-coherent feedback — the signature of a network entering collapse.


Zogg Fire — September 27, 2020, Shasta County

Four people killed. 204 structures destroyed. PG&E equipment caused the ignition. Backtested on a 33-node Shasta County CEC transmission topology — the sparsest grid in the dataset. DON substrate evolution amplified entropy to 0.5819 peak (+17.2% above raw) — the strongest DON boost in the 14-fire corpus. Nearly the entire 33-node network crossed θc simultaneously.

Zogg Fire Day — DON-Evolved Entropy

33-node Shasta County grid. Peak DON entropy 0.5819. Threshold = 0.42. 21 of 24 hours above threshold. 16 hours flagged as high risk.

System Response — Zogg Fire

Near-total network collapse: 48.5% of edges above θc in raw measurement.
21
of 24 hours above θc
16
Hours flagged as high-risk
33
Nodes — sparsest grid tested

The Zogg Fire represents near-total network collapse — so many edges are simultaneously above θc that the DON beam search has limited safe alternatives. 16 of 24 hours were flagged as high-risk. On sparse grids, the system signals that de-energization may be the only safe response. The +17.2% DON boost (raw 0.4963 → DON 0.5819) reflects the entire network entering a coherent collapse state.


Kincade Fire — October 23, 2019, Sonoma County

374 structures destroyed. PG&E transmission equipment caused the ignition near Geyserville. Backtested on a 261-node Sonoma County CEC topology — the densest grid in the dataset. DON-evolved entropy above θc for 15 of 24 hours with first detection at 9:00 AM. The dense grid provided many alternative paths for the DON beam search.

Kincade Fire Day — DON-Evolved Entropy

261-node Sonoma County grid. Peak DON entropy 0.5242. Threshold = 0.42. 15 of 24 hours above threshold. Dense topology enables effective rerouting.

System Response — Kincade Fire

Dense topology enables effective DON beam search across many candidate paths.
15
of 24 hours above θc
14h
DON maturity sustained
261
Nodes — densest grid tested

The 261-node Sonoma County grid has 546 edges — many possible paths between any two points. The DON beam search evaluates K candidate paths per routing cycle, selecting minimum-action routes through the network. The dense topology provides enough alternatives that elevated entropy is navigated around rather than through.


Dixie Fire — July 13, 2021, Butte County

The largest single fire in California history. 1,329 structures destroyed. PG&E equipment caused the ignition. Backtested on the same 104-node Butte County CEC topology as the Camp Fire — sharing the same transmission corridors. A summer fire with lower winds than the autumn Diablo wind events. DON-evolved peak entropy 0.5236 with 13 hours above θc.

Dixie Fire Day — DON-Evolved Entropy

104-node Butte County grid. Peak DON entropy 0.5236. Threshold = 0.42. 13 of 24 hours above threshold. Summer fire with clear diurnal profile.
Cross-fire consistency

System Response — Dixie Fire

Same grid as Camp Fire. Three years later. DON beam search finds the same structural vulnerabilities.
13
of 24 hours above θc
11h
DON maturity sustained
104
Same Butte County grid as Camp Fire

The Camp Fire (November 2018) and Dixie Fire (July 2021) occurred on the same 104-node Butte County grid. Running independently — different weather data, different date, different conditions — the DON beam search routes around the same structural vulnerabilities both times. This is physics finding the same collapse-susceptible corridors in the grid, not a trained model remembering a pattern.


Collapse Morphology

The Camp Fire's unique signature

Most fires show diurnal entropy oscillation — rising with afternoon heat and wind, falling overnight. The Camp Fire did not. Its entropy trajectory was flat, sustained at elevated levels around the clock. This is the signature of a system trapped at the bifurcation boundary — decoupled from the solar cycle that normally drives weather patterns.

Camp Fire — Bifurcation Plateau

0.953
Flatness
0.04
CV
5
Crossings
18/24
Hours above

Flatness = avg/peak ratio. A value of 0.953 means the average entropy was 95.3% of the peak — virtually no variation. The coefficient of variation (CV) of 0.04 is the lowest in the 14-fire corpus. The Camp Fire is the only fire classified as a PLATEAU morphology. All other fires show MULTI-PEAK (diurnal oscillation) or SPIKE (rapid onset/recovery) patterns.


Severity Indicator

Entropy dynamics correlate with fire severity

Peak entropy does not predict fire severity (r = 0.009 — effectively zero correlation). But LDEE, a composite measure of how entropy crosses and persists at the bifurcation boundary, shows strong correlation with fire outcomes (r = 0.923, p = 0.000002). This is not a death predictor — it is a severity indicator derived from collapse dynamics.

What LDEE measures

LDEE = crossings × persistence × dash_hours. It captures three physical properties: how many times entropy crosses the bifurcation boundary (crossings), what fraction of above-threshold time is contiguous (persistence), and total hours above threshold (dash_hours). A high LDEE means the system is repeatedly entering and sustaining collapse conditions — not just briefly spiking above the boundary.

LDEE by fire — sorted by severity indicator

Camp Fire 85
56.2
Redwood Valley 9
16.7
Atlas Fire 6
9.8
Cherokee Fire 0
7.3
Zogg Fire 4
6.4
Dixie Fire 1
5.2
Kincade Fire 0
3.1
Cascade Fire 4
1.3
Tubbs Fire 22
0.7
Nuns Fire 3
0.7

Left number: fatalities. Right number: LDEE value. The Camp Fire is a massive outlier (LDEE = 56.2) — it crossed the bifurcation boundary 5 times, with 62.5% persistence, sustaining 18 hours above threshold. Four fires with zero LDEE (Valley, Butte 2015, Jerusalem, Rocky) are omitted — they never crossed the boundary in raw analysis.

Tubbs Fire anomaly

The Tubbs Fire (22 deaths, LDEE = 0.7) is the notable outlier in the correlation. Its high death toll resulted from population density and evacuation dynamics in the Santa Rosa urban interface — not from sustained grid collapse conditions. LDEE measures grid field dynamics, not population exposure. This distinction is important: the indicator reflects collapse severity, not community vulnerability.


Structures destroyed

Major California wildfires — utility-investigated or utility-caused. Sources: Cal Fire incident reports.

Lives lost

Major California wildfires — utility-investigated or utility-caused. Sources: Cal Fire incident reports.

Data Sources & Methodology

Grid topology: Three regional CEC transmission networks — Butte County (104 nodes, 224 edges), Shasta County (33 nodes, 68 edges), and Sonoma County (261 nodes, 546 edges) — derived from California Energy Commission GIS public data (6,839 features). Weather: ERA5 reanalysis hourly data fetched at each node's geographic coordinates via Open-Meteo archive API. Load: CAISO historical demand data. Load: CAISO OASIS API hourly data (2009+), scaled to local grid capacity. Detection threshold θc = 0.42 is the DON Theory spectral bifurcation boundary, empirically validated across 14 fires. DON substrate evolution amplifies spatially-coherent entropy. Mean boost: +14.1% across all 14 fires. False positive rate: 0% (0/300 topology-days) with unified detection gate, tested across 100 non-fire reference dates on all three topologies.

CEC Transmission Lines GIS Dataset • Open-Meteo ERA5 Archive • CAISO Open Access • Backtest run: February 2026

Firebreak is ready for a pilot deployment

The backtest demonstrates detection capability on historical data. The next step is a controlled pilot on a live grid segment to validate performance under operational conditions. We are seeking utility and regulatory partners for that pilot.

Discuss a Pilot
DON Systems LLC  |  partnership@donsystems.com