# Coupled energy enterprise: reviewed evidence notes

Prepared 27 October 2026 for the Laplace investigation page. This page compares recorded synthetic outputs; it does not run a solver in the browser.

## Displayed files

- `replay.json`: selected three-policy counterfactual replay, event chronology, corrected four-arm held-out comparison and interpretation boundaries.
- `coupled-enterprise-model.png`: unmodified source `fig_dependency_graph.png`.
- `heldout-policy-outcomes.png`: unmodified source `fig_headline.png`.

Interpretation follows `counterfactual_traces.json`, `key_results.json`, `heldout_comparison_v2.csv`, `heldout_comparison.NOTE.txt`, `CORRECTIONS.md`, `ablation.csv`, `validation_report.json`, `model_inventory_validated.csv`, `interface_contracts.csv` and `assumption_register.csv`. The corrected held-out definitions and written corrections supersede earlier timing summaries.

## Synthetic scope and complexity

The model is a fictional two-region utility around the IEEE 39-bus reference network, with 46 branches, 12 mixed-fidelity components and 29 declared cross-domain contracts. It connects electrical supply with cooling, gas, communications, digital services, crews and operator response. No measured company data or field calibration is supplied. Injected cyber effects do not constitute an intrusion or attacker model.

Twenty-three supplied component, interface and reproducibility checks passed. They check specific properties. Six components, including policy and digital layers, lack dedicated external or analytic reference checks. The validation run publishes 22 signals and reads three through the bus; it does not demonstrate that all 29 declared contracts were exercised.

## Held-out comparison

Each of 4,948 feasible held-out scenarios is compared under four policy arms using the same disturbances, environment and seeds over six simulated hours. Of 5,000 sampled scenarios, 52 were excluded by the pre-disturbance feasibility screen before scoring. A valid starting dispatch is not a guarantee of subsequent safety.

The metric is unserved energy as a percentage of total energy demand. Lower is better.

| Policy | Mean unserved energy | Supplied 95% interval |
|---|---:|---:|
| No operator action | 12.4059% | 11.9312%–12.8827% |
| Coordinated response | 12.7613% | 12.2678%–13.2517% |
| Rule-based baseline | 16.0383% | 15.4473%–16.6450% |
| Per-subsystem response | 16.3040% | 15.7708%–16.8316% |

The supplied paired penalty versus no operator action is +3.6324 percentage points for baseline (95% interval +3.2773 to +3.9933) and +0.3554 points for coordination (+0.1902 to +0.5140). Coordination improves on baseline but does not beat no operator action on this held-out mean. The per-subsystem point estimate is higher than baseline, but their difference is not statistically resolved by the supplied paired interval.

No operator action removes only the operator policy layer. Automatic frequency control and five-minute security-constrained redispatch remain active. This control arm is not operational advice.

## Selected replay and causal boundary

The selected scenario has the following unserved-energy fractions:

| Policy | Unserved energy | Unserved energy / demand |
|---|---:|---:|
| Rule-based baseline | 20,686.2393 MWh | 91.0058% |
| Coordinated response | 2,302.7350 MWh | 10.1305% |
| No operator action | 2,880.8501 MWh | 12.6738% |

All three branches have 22,730.6837 MWh total demand. This severe selected replay favors coordination more strongly than the held-out population; it must not be presented as a typical scenario.

The chronology prevents a cyber-specific causal claim:

1. A physical line trip is scheduled at 1,649.4187 seconds; the event log records it at 1,650 seconds.
2. Baseline orders another circuit open at 1,695 seconds; the switch execution is logged at 1,696 seconds.
3. The first injected cyber effect starts at 11,590.0498 seconds.

The operator command follows the physical trip by about 46 seconds and precedes the cyber effect by about 2.75 hours. The principal cascade begins before that cyber effect. The replay supports a modeled response-doctrine comparison under combined stress, not a claim that cyber caused the main outage. Several actions differ between policy arms, so the comparison does not isolate one command's causal effect.

## Mechanisms that did and did not matter

Baseline opens a circuit in 775 of 4,948 held-out scenarios. Its excess unserved energy versus no operator action averages +18.49 percentage points there, compared with +0.87 points in the other 4,173 scenarios. This is a subset association supported by the trace, not an isolated randomized command effect.

Exploratory ablations use 250 selected worst development cases, not the held-out population. Removing thermal-to-power coupling reduces mean unserved energy by 5.67 points. Removing either gas-to-power or power-to-gas coupling changes it by exactly zero. The expected gas-electric mechanism was inactive in that selected subset. Communications isolation was also not the dominant damaging reflex in the inspected evidence.

## Figure populations and remaining limits

The original policy-outcome figure has panels with different populations: all held-out policy outcomes, a circuit-opening subset split, and ablations on selected worst development cases. Their denominators must remain separate.

Earlier detection and restoration definitions were corrected. Their corrected means apply only to cases for which those times are defined, so they are not featured as typical recovery times. Full-campaign aggregate counts are not used on this page because the supplied raw-wave accounting and final prose have unresolved inclusion differences.

The model demonstrates consequences within a synthetic enterprise. It does not establish real utility resilience, operational safety, field performance or a general recommendation to avoid operator action.