Nuclear probabilistic risk assessments predicted core-damage frequencies well below the observed historical accident rate

The claim rests on two required premises and faces one substantive counterargument.

Predicted side: the record of published estimates is well documented. WASH-1400 (1975) estimated core melt at roughly 5×10⁻⁵ per reactor-year; the NRC requires designs to meet a 1×10⁻⁴ core-damage frequency; early-2000s US fleet estimates ranged 2-5×10⁻⁵ (visualizingenergy.org/what-do-these-charts-say-about-the-safety-of-nuclear-reactors/); TEPCO's PRAs for Fukushima Daiichi predicted about 1×10⁻⁶ per reactor-year at power (www.ncbi.nlm.nih.gov/books/NBK253921/). The subclaim that PRAs typically estimated frequencies of 1e-4 to 1e-5 per reactor-year or lower↗︎ is effectively uncontested.

Observed side: Three Mile Island (1979), Chernobyl (1986), and Fukushima (2011, three cores) are the undisputed core, with broader counts reaching ~10-11 events when partial melts (Saint-Laurent, Greifswald, Fermi 1, Lucens, etc.) are included. Over ~14,000-18,000 cumulative reactor-years this yields ~2×10⁻⁴ to ~7×10⁻⁴ per reactor-year (www.sciencedirect.com/science/article/abs/pii/S0925753513002816, arxiv.org/pdf/1207.5674). The order of magnitude is robust to counting choices, so the observed-rate subclaim↗︎ holds on every credible tally.

The counterargument, that the record is dominated by designs and eras outside PRA scope↗︎, is partly true: Chernobyl's RBMK had no Western-style PRA, most events predate 1990, and post-1986 event rates dropped and stabilized (arxiv.org/pdf/1504.02380). But it cannot carry a full rebuttal, because Fukushima involved reactors with plant-specific PRAs predicting ~1×10⁻⁶ per reactor-year, and the dominant failure there (common-cause external flooding) was exactly the kind of contribution the PRAs under-weighted. Even the most restrictive comparison, Western light-water reactors with PRAs against their own operating history, shows observed frequency above prediction.

Verdict: supported rather than verified, because the size of the gap depends on contestable reference-class and event-counting choices, and a fully rigorous like-for-like actuarial study across the fleet was not within this pass. What would change the conclusion: a demonstration that, restricted to reactors actually covered by published PRAs and to the operating periods those PRAs describe, the observed core-damage rate falls within the predicted range. Post-Fukushima operating experience accumulating without further core damage will also gradually pull the observed rate down toward the predicted range.