A cooling-demand refrigerant-bank prior does not improve population for HFC-134a inversions (a negative result)

A negative result, published so the next person does not have to re-discover it: for spatial disaggregation of European HFC-134a, a refrigerant-bank prior weighted by population × climatological cooling-degree-days (CDD) is worse than the plain population proxy already used by EDGAR. Pooled spatial correlation against the ICOS PARIS 2020 HFC-134a inversion posterior (6-member ensemble, exact-2020 fields): BANK 0.237 vs population 0.616 vs EDGAR 0.655; the bank prior beats population in 0 of 11 scoreable countries (and is unscoreable — identically zero — in 4 more, an artifact of monthly-climatology CDD evaluating to zero across maritime/Nordic Europe; excluding those countries entirely the bank prior still loses 0.262 to 0.579). HFC-134a is a diffuse, population-coupled source: refrigerant banks co-locate with people, so the cooling-demand re-weighting tested here only degrades the population baseline. Companion (the structural opposite, where the facility prior wins): the smelter-resolved CF4 prior, DOI 10.5281/zenodo.20617486. Scope, honestly: one proxy was tested (the multiplicative climatological-CDD extreme of pop × (a + b·CDD)); daily-resolution CDD, mixed weightings, and non-climate stock proxies were not run and the negative is not claimed for them. The posterior is prior-anchored (it was run from an EDGAR/population-shaped prior; it correlates with its own prior at 0.687 pooled, Greece 0.999), which limits the test's power — the negative rests on the candidate producing zero positive evidence anywhere, including the least-anchored countries, with the companion CF4 test as the positive control showing the harness detects real prior improvements. A "what would overturn this" list is in the validation report. Version 2 (2026-06-12), adversarial-audit revision: (1) a year-selection bug was found and fixed — v1 selected the posterior year with nearest-neighbour matching, which silently returned the 2019 field for 4 of 6 ensemble members, so v1's published numbers (0.270 / 0.634 / 0.671) were computed from a 4×2019 + 2×2020 mixture; all numbers are recomputed from exact-2020 fields (the qualitative verdict is unchanged) and the audit script ships in this version. (2) The v1 claim that the losing margin was one "the prior bias cannot explain" is withdrawn; the prior-anchoring column is now published. (3) "0 of 16 scored countries" corrected to 0 of 11 scoreable. (4) The northern-zero mechanism is correctly attributed to the monthly-climatology CDD construction rather than to the climate. (5) Formal data citations added (ICOS/Ganesan et al. 2026; EDGAR/Crippa et al. 2024; GHS-POP/Schiavina et al. 2023; WorldClim/Fick & Hijmans 2017). (6) License correction: the WorldClim-derived files (cdd_europe.nc, prior_hfc134a_europe_bank.nc) are removed from this version because WorldClim 2.1 is free for academic/non-commercial use and may not be redistributed without permission, which is incompatible with this deposit's CC-BY-4.0 license; regenerate them locally via src/build_bank_prior.py (one command; inputs are public). The population-only prior (GHSL-derived, CC-BY) remains. Code also at: https://github.com/zachdissington/climate-solutions-research-kit (priors/hfc134a-refrigerant-bank-prior/). Author ORCID: 0009-0007-0196-1371.