O-14 - The RHIC strangelet safety analysis ran five years before its anthropic-selection flaw was found and corrected
The corrected analysis (anthropic bias removed, using Jaffe et al. 2000 inputs plus a planetary-formation-rate model) still bounded accelerator risk below 1e-12/yr, but the episode is a documented instance of a serious flaw persisting undetected in exactly the class of physics safety argument at issue for the LHC. The paper also notes Kent (2004) found mistakes in how risk probabilities were stated across versions of the Dar et al. paper and the Brookhaven report.
Link to original
Why this is evidence
The most direct discriminator in the corpus: a published accelerator-safety argument in precisely the class at issue (cosmic-ray/Earth-survival bounds) contained a serious unnoticed flaw — ignored anthropic observation selection — that survived peer review and years of operation before Tegmark & Bostrom corrected it. H-11 predicts flaws of this kind occur at a non-negligible rate in exactly these arguments; H-46, which holds that this argument class is robust enough for its conditional bounds to carry through undiluted, makes a documented in-class failure distinctly less expected. (That the corrected analysis still bounded the risk low is likelihood-relevant detail for the pricing step, not a reason to withhold the edge.)
Likelihood
# E-27 (HC-14) — lone edge, one observation O-14 "the RHIC strangelet safety analysis ran ~5 yr before its
# anthropic-selection flaw was found & corrected". Members in HC-14.hypotheses order [H-11, H-46];
# exhaustive_by_construction, NO residual. Price ONLY the INCREMENTAL discrimination of this one same-class
# datum: the prior already rests on E-24 (general scientific error rates ~1e-3), E-26 (Castle Bravo model
# failure) and A-2 (layering). What O-14 adds beyond those outside-view facts is a precedent inside the EXACT
# genre at issue — a pivotal cosmic-ray/Earth-survival accelerator-safety argument that carried a serious
# unnoticed flaw through ~5 yr of operation. Anchored on H-11 = 1 (rule 7).
lik_rhic_H11 = 1.0 # anchor: H-11 (argument-failure dominates the actionable P(X)) predicts flaws in exactly
# these arguments at a non-negligible rate — a documented in-class instance is just what it expects
lik_rhic_H46 = 0.55 # 0.55x as expected under H-46 (arguments robust, conditional bounds carry through
# undiluted): a serious in-class flaw persisting ~5 yr cuts against robustness. Not lower
# because only the same-class SPECIFICITY is new here (general error rates already priced
# into the prior via E-24/E-26), it is n=1, and it was self-correcting — the corrected
# analysis still bounded risk <1e-12/yr, the kind of layered self-repair H-46 leans on
t_rhic = 0.75 # cap = trust_score(S-44, Ord-Hillerbrand-Sandberg) = 0.8; docked slightly: the timeline
# is solid historical fact, but the "serious flaw" severity is the source's interpretive
# judgment (the corrected bound stayed benign), so how much weight the episode carries is contestable
evidence("HC-14", ["O-14"], [lik_rhic_H11, lik_rhic_H46], t=t_rhic)