The Guarino Infrastructure Dependency Metric, Paper II: Plant-Level Buffer Groups and the Failure Sequence of the Texas Chain

Revision 2, 20 September 2026. Paper II of the Guarino Infrastructure Dependency Metric. Continues Paper I (10.5281/zenodo.22866438, revision 2). Revised after external review: prediction protocol with four hypotheses stated before the results, evidence labelled as observed, attributed or inferred, threshold and join sensitivity, Kaplan-Meier and Cox survival analysis, the usable buffer definition, and narrowed claims. The deposit carries the manuscript, LaTeX source, the four public source files as downloaded, two analysis scripts that reproduce every table and statistic, all intermediate tables, figures, a manifest and checksums. Paper I of this series derived the buffer group ΠB = Ta/Tx on every edge of an infrastructure dependency graph and reported the February 2021 Texas chain edge by edge as bound or held. This paper operationalises the group against four public datasets: ERCOT's unit-level outage record for 10 to 19 February 2021, the EIA-860 plant inventory for 2020, the EIA-923 fuel stock and consumption schedules for 2021, and ERCOT's fuel mix report at fifteen-minute resolution. Four hypotheses are stated before the results and each is tested separately, with the evidence for every claim labelled as observed, attributed or inferred. First, defining failure as a unit's first loss of at least half its seasonal capacity, the capacity-weighted median onset by fuel type was wind at 29.5 hours from midnight on 10 February, coal at 110.5, gas at 122.1 and nuclear at 125.5, against a load shed order at 121.3. Wind first and nuclear last are what Ta requires, the nuclear endpoint resting on a single unit; the gas-coal order is threshold-sensitive, matching the prediction at five of six loss thresholds and inverting at the 50 percent threshold on eleven coal units. ERCOT's aggregate attribution puts the weather edge at 27,567 MW and the fuel edge at 6,130 MW at peak unavailability, and the gas fuel edge appears as a 15 February spike holding 35 to 47 percent of gas major losses at every threshold. Second, among 193 failed gas units joined to their EIA-860 records, units at plants with two or more pipeline connections failed later: log-rank p = 0.006, Cox hazard ratio 0.73 per connection (95 percent CI 0.62 to 0.87) controlling for unit size and oil switching, robust to three join rules. Their higher availability over the emergency window is not robust: the bootstrap interval on the difference includes zero and the strict join loses significance. Oil-switching capability shows no onset effect. The gas result is associational. Third, end-of-January coal stock over burn rate gives ΠB of 1.8 to 2.0 at the thinnest pile and above 2.3 elsewhere under January, February and maximum-month burn rates, at a fleet that lost 5,117 MW at peak: inventory exhaustion was not the limiting mechanism, and the paper defines the usable buffer Sb,u that was. Fourth, the fuel mix report shows the available fleet running at ΠP of at least 0.80 for gas and 0.98 for nuclear while ΠA fell, the signature Paper I assigns to buffer failure. The validated claim is that the record is consistent with the buffer group's edge-level regimes; unit-level causal attribution, and the relation ΠA = min(1, ΠB) at a unit, remain untested by these data.