Utilities

01 The Decision

UTILITY-CAUSED WILDFIRE FOLLOWS A TEXTBOOK DIRECTED ACYCLIC GRAPH. Equipment condition determines whether a conductor fails during high winds. Wind severity and vegetation proximity determine whether a failed conductor produces ignition. Ignition determines whether a catastrophic fire results. Under California’s inverse condemnation doctrine, liability is the downstream consequence of the entire causal chain, not of negligence, but of causation. A risk score gives the board a number. The causal model gives the board a mechanism: which equipment decision, on which span, would have changed which outcome, by how much.

The model also reveals the central confound in observational PSPS analysis. Utilities de-energise lines during Red Flag Warning conditions, the same conditions that drive high wind severity. Any analysis that treats PSPS status as observed rather than intervened on will systematically underestimate the protective effect of de-energisation, because the observed data conflates the intervention with the weather that triggered it. Only the do() operator severs this back-door path.

02 The solution

Black check mark = observed evidence (a fact entered as a filter). Red check mark = do() intervention (a decision applied, severing the influence of parents).

The solution was to model the relationships between the variables that determine ignition risk, and to be explicit about which variables cause which. We recognised, for example, that PSPS de-energisation decisions are not independent of weather conditions, utilities de-energise most aggressively when Red Flag Warnings are active, which are the same conditions under which ignition risk is highest. That correlation means observational data understates how much de-energisation actually reduces risk, because it is always deployed against the hardest conditions. Equipment age and maintenance history both influence ignition probability, but through different mechanisms: one determines baseline failure rate, the other determines whether a known deficiency was addressed before a high-risk period. When you observe a variable in this model, you are effectively filtering the data to cases where that variable takes a particular value, and that filter ripples through the model, shifting related variables up and down accordingly. When you intervene on a variable, forcing it to a value regardless of what caused it, you break that ripple effect and get a cleaner answer: not what circuits that look like this tend to produce, but what would have happened if this specific condition had been different When you abduct, you extract a particular case from the averages, locking in its idiosyncratic circumstances before asking what would have happened if one or more things had been different.

03 What the solution answers

Rung 2 tells you what replacing equipment does on average. Courts ask a narrower question: given this specific circuit, this specific weather, this specific incident history, what would have happened if that fitting had been in good condition?

Lines are de-energised during the same Red Flag events that drive high ignition risk. Setting PSPS = De-energised as an intervention severs that back-door path.

Enter observed evidence; read which upstream states become more probable. Identifies circuits to inspect. Does not assert causation, that is Rung 3.

04 Inside the Model

A language model can speak fluently about any domain. It cannot know one. The .bayes file is the knowledge the LLM is missing: a causal map of the domain, auditable, versioned, and wrong in specific correctable ways.

Rung 1, Diagnostic: where does ignition risk concentrate?

Prior obs(Liability = High) + obs(Wind = Extreme) + obs(Incident = Severe)

Prior, no evidence set

Baseline. EC: 49.5% Good / 32.1% Degraded / 18.5% Failed. Ignition: 73.0% No / 17.2% Probable / 9.8% Certain. Liability: 75.6% Low / 12.9% Med / 11.5% High.

Rung 2, Intervention: do() vs obs() on PSPS

do(PSPS = De-energised) obs(PSPS = De-energised)

do(PSPS = De-energised), causal estimate

Back-door severed. Red Flag stays at prior. Ignition: 97.4% No. True causal effect isolated.

Rung 3, Counterfactual: would the fitting replacement have prevented the fire?

Step 1, Abduction Step 2, do(EC = Good)

Step 1, Abduction: factual conditions entered

U_EC shifts from 50/50 to 37.4% Fav / 52.6% Adv. This circuit's background anchored. Ready for counterfactual.

05 Just Ask

The model is a file. Any capable LLM can load its XML and answer wildfire liability and PSPS questions in plain English.

Same model, three rungs. The audit trail is the .bayes file, not the risk score.

06 The Engagement

If a fire starts from utility equipment tomorrow, the question will not be whether wildfire was on the risk register. It will be whether the investment decisions were the best use of available resources given what was knowable. The causal model provides that answer, before the fire, not after.

• Pre-litigation defence. A counterfactual probability with explicit assumptions survives cross-examination. A risk score does not.
• PSPS cost-effectiveness. Tightening thresholds requires knowing the true causal reduction per event, not the observational estimate confounded by the weather that triggers PSPS.
• Capital allocation. The diagnostic identifies which circuits, under which conditions, are most consistent with a high-liability outcome.

This case study is a composite drawn from published literature and practice. Specific figures are representative. No individual organisation or engagement is described.