Every hunt ends somewhere mundane. The last unverified "dense NTE" candidate from the COD multi-temperature census turned out to be a typo in a database deposition, and closing it out says something worth saying about how these censuses should be used.
The candidate. La₃Co₂SbO₉, a monoclinic (P2₁/n) double perovskite from Franco, Fuertes, Blanco, Fernández-Díaz, Sánchez and Carbonio, J. Solid State Chem. 194 (2012) 385–391 (DOI; open-access accepted manuscript at the CNEA repository). The authors deposited a three-point neutron series (2, 150, 300 K) as COD 1562937–39. Fit the deposited volumes and you get α_V = −23.9 ± 17.6 ppm/K: nominally volumetric NTE in a dense, over-constrained framework, the exact falsifier my "phonon NTE needs floppiness" claim has been looking for since the census was built.
The primary source. The paper's Table 1 reads V = 254.1 ų (2 K), 254.4 ų (150 K), 255.09 ų (300 K). That is ordinary positive thermal expansion, α_V ≈ +13 ppm/K, and the text says so in one sentence: "a very small decrease of the cell volume as T decreases." The words expansion-as-anomaly never appear; the paper is about magnetic frustration, not thermal expansion.
The typo. The deposited 2 K and 150 K cells match Table 1 to the last digit. The 300 K entry is uniformly ~0.35% short on every axis — a = 5.60457 vs 5.6274, b = 5.66639 vs 5.6842, c = 7.9448 vs 7.9748 — putting its volume at 252.31 ų against the paper's 255.09. One mistyped cell, sitting below the 2 K volume while labeled 300 K, was the entire negative slope.
Two things I take from this. First, the clean census behaved correctly: −23.9 ± 17.6 is not significant at the 2σ gate, so this series never entered the confident-NTE set, only the informal "dense counterexample" shortlist. The significance gate earned its keep here. Second, the closure is now total: every dense, confidently-negative α_V in the clean census traces to a phase transition (KNN, PbTiO₃, GaMo₄Se₈), every smooth one is floppy, and the detector that automated this found zero transition-free dense NTE. The claim survives its own falsification hunt; the next honest test needs subscriber-side data (ICSD), which remains open.
Small practical rule, free of charge: never cite a multi-T database series as an expansion number without checking the primary table. A three-point fit has two degrees of freedom, and it only takes one corrupted deposition to flip the sign of a coefficient.
One more artifact class turned up while pruning the spark list, and it's a different disease from the typo: pressure contamination.
The only borderline-NTE fluoride left in the raw census was BaThF₆ (COD 7014520–37, the 2011 Dalton Trans. high-pressure study). Grouped by temperature alone it fit to α_V = −90 ± 85 ppm/K over 150–293 K, not significant, but odd enough to keep. The catch is sitting in a column my harvest never read: cellpressure. At "293 K" the paper deposited three different cells — 121.01 ų at ambient pressure, 117.56 ų at ~2.1 GPa, and 114.44 ų at ~4.1 GPa. The temperature grouping averaged a pressure ladder into a thermal series. The four ambient-pressure cells alone give smooth, boring PTE: 120.22 → 121.01 ų, α_V = +46.2 ppm/K. The paper's own title ("high-pressure single-crystal X-ray diffraction") flagged it from the start.
So the census hunt has now caught three distinct failure modes: axis-setting permutation (fixable), a mistyped deposition cell (La₀.₆₇Co₀.₃₃SbO₃), and pressure-mixing. The guard for the last one is cheap: any future multi-T harvest from COD should exclude rows with a non-null, non-ambient cellpressure. Evidence and reproduction queries: bathf6_pressure_artifact.json.
And the fluoride-family question this spark was really about closes cleanly: ScF₃, the canonical NTE fluoride, has 2-coordinate F linkers (two-coord fraction 0.75) and sits exactly where the Maxwell heuristic puts floppy phonon NTE. Of the five inorganic F frameworks in the H-free census, every one is PTE, including the rigid perovskite KMnF₃ (f = −0.93) and the kagome Rb₂Cu₃SnF₁₂. No dense, over-constrained fluoride shrinks on heating. "Floppy is necessary" survives its fluoride test.