Measurement Blind Spots · supplied 1 November 2026
30 refined leads
Analytical derivations and simulations are source-reported, not independently reproduced during publication. No published dataset was directly reanalysed. Source readiness/novelty scores are author assessments, not validation or novelty clearance. Original claims, corrections and negative results are retained.
Source-selected development, not an independently verified evidence tier. The linked original identifier remains the hypothesis’s identity.
R01 · Finite-window decay exponents are set by the window, and the offset is recoverable from curvature
Refines X01. Sources: investigation I1; baseline-subtraction practice.
Claim. For a decay y(t) = At^(−α) + c with an unsubtracted offset c, the local log–log slope is exactly s(x) = −α/(1 + Be^{αx}) with B = c/A and x = ln t. A straight-line fit over a window therefore returns α_fit = α·⟨logistic⟩, where the average is taken with the least-squares weight.
What is an identity. The logistic form of s(x) is an identity — it follows from differentiating ln y. So is the result that a window centred on the crossover t_c = (A/c)^{1/α} returns exactly α_fit = α/2. This is not a discovery about nature; it is a property of the estimator, and it should be quoted as such.
What is a genuine prediction. Three things are not identities. (i) The bias is bounded: for a window entirely below the crossover the bias is <1% even at R = 1 decade, so the artefact cannot explain a literature-wide discrepancy unless the crossover lies inside the fitted range. (ii) The offset is identifiable from curvature: a three-parameter fit recovers α with zero bias (Monte-Carlo mean α_fit/α = 1.000 ± 0.011 at R = 3 decades, 2% noise) even when the two-parameter fit is 50% low. (iii) The Cramér–Rao bound diverges when the window sits below the crossover, so there is a hard limit on what any offset-blind analysis can claim.
Discriminating observation. Re-fitting a published decay with a free offset should shift the exponent, and the shift should be predicted by the fitted crossover. The diagnostic needs only the published data points.
Refutation. The hypothesis is refuted if published exponents are unchanged when a free offset is allowed and the log–log curvature is consistent with zero at the stated noise. If exponents move but the movement is not predicted by t_c, the mechanism is wrong even though the bias is real.
Limits. The analysis assumes additive, constant-variance noise in y. With multiplicative noise the bias changes sign structure; that case is not covered. A drifting (rather than constant) offset produces a time-dependent bias that the three-parameter fit cannot absorb — a genuine degeneracy, not an artefact of the fit.
Practical use. A two-line check on any reported power law: fit with and without a free offset, and report both. If they differ by more than the stated uncertainty, the exponent is a property of the window.
R02 · Kramers–Kronig consistency is not a linearity test
Refines C01. Sources: investigation I3; describing-function theory; arXiv:2605.08251 for the extrapolation analogue.
Claim. A Butler–Volmer interface shunted by a linear double-layer capacitance has a small-signal describing function that is exactly a Debye element with an amplitude-dependent conductance. It therefore satisfies Kramers–Kronig identically at every amplitude, while the charge-transfer resistance it returns is biased low by x/(2I₁(x)) with x = A/b.
What is an identity. With no series resistance,
Z₁(ω) = 1/(G_nl(A) + jωC_dl) exactly, and the Bessel result G_app/G_true
= 2I₁(x)/x is exact. Harmonic balance reproduces both to 1e-17
(validation in I3_kk_blindspot.py). None of this is new
physics; it is the standard describing function of a symmetric
exponential nonlinearity.
What is a genuine prediction. The combination — that the standard validity test cannot detect the bias — is falsifiable and non-obvious. Numerically, the finite-amplitude spectrum stays a Debye to 1e-7 relative even at 100 mV, while R_ct falls to 0.644 of its low-amplitude value (x = 1.95). A Lin–KK residual does not track the bias (0.170 at 2 mV, 0.048 at 100 mV): the test moves the wrong way.
Quantitative. R_ct/R_ct,true = 1.000 at 2 mV, 0.996 at 10 mV, 0.891 at 50 mV, 0.644 at 100 mV for b = 51.4 mV. The third-harmonic ratio I₃/I₁ measures x to 0.2% and converts directly into the correction.
Discriminating observation. On a reversible outer-sphere couple, R_ct should fall quadratically with amplitude with the predicted coefficient, while I₃/I₁ rises as x²/48.
Refutation. If R_ct is amplitude-independent to 1% across 2–100 mV on a system with a known Butler–Volmer response, the mechanism is absent. If R_ct varies with amplitude but I₃/I₁ does not track it, the mapping is wrong.
Limits. The mapping x → bias assumes a symmetric barrier and no mass transport. With asymmetric transfer coefficients a DC bias generates a second harmonic that the symmetric model does not predict; with a Warburg element in series the harmonic current is loaded differently and the naive correction will over- or under-shoot. Those cases are stated, not solved.
Practical use. Report R_ct at two amplitudes, or report I₃/I₁ alongside every impedance spectrum. Either costs one extra measurement and bounds a systematic that no KK test can see.
R03 · In-plane thermal conductivity from optical thermometry is a function of spot size, and in one limit it is not a material property at all
Refines M01. Sources: investigation I2; steady-state Raman thermometry modelling.
Claim. For a Gaussian source on a sheet with substrate conductance G, ΔT = (P/4πκd)·e^λE₁(λ) with λ = Gs²/(2κd). The standard fit returns κ_app = κ/[e^λE₁(λ)], which crosses unity at λ = 1 and asymptotes to κ_app ≈ Gs²/(2d).
What is an identity. The closed form is exact for the stated model — verified against direct Hankel quadrature to 2e-13. The asymptotic κ_app → Gs²/(2d) is a limit, not a new law.
What is a genuine prediction. The crossover λ = 1 corresponds to a spot radius s* = √(2κd/G): 0.58 µm for κ = 50 W/mK, d = 100 nm, G = 3×10⁷ W/m²K. That is squarely inside the operating range of Raman thermometers, so the bias changes sign across a typical experimental campaign. The prediction is a collapse: reported κ for one material should scale onto one curve when plotted against λ computed from each study’s spot radius, and should scale as s² for large spots.
Quantitative. For plausible (s, G) ranges the model produces a 136-fold spread in reported κ for a single true value. At λ = 1 the bias is +68%.
Discriminating observation. Reanalysis of a published κ compilation with spot radius and G metadata. The falsifiable signature is a slope of 2 on a log κ_app vs log s plot at large s.
Refutation. κ_app uncorrelated with spot radius across a compilation spanning 0.2–3 µm refutes the mechanism. So does a κ_app that is independent of the substrate.
Limits. The model is isotropic, steady-state, optically thin, and treats the substrate as a single conductance. Anisotropy, finite substrate thickness, and the transducer’s own boundary all add terms that are not modelled here; the 136-fold figure is an upper bound for the stated parameter range, not a prediction for any named material.
Practical use. Report the spot radius and the substrate conductance with every 2D κ measurement; fit λ rather than κ. Two spot sizes break the κ–G degeneracy — more laser powers do not (Cramér–Rao 0.037 at a 1:3 spot ratio vs rank-deficient at one spot).
R04 · Negative result: stray light does not explain OD600 compression in the exponential phase
Refines B01. Source: R1_refinement_numbers.py,
section A.
Original claim. Plate-reader optical density compresses above ~0.3–0.6 because of detector saturation, suppressing fitted growth rates.
What the calculation shows. Modelling the reader as an additive stray-light fraction S, the fitted growth rate is suppressed by 0.1% for S = 10⁻³ over the window OD 0.2–1.0, and by 1.1% for S = 10⁻². The effect only becomes large where assays rarely operate: −8% for S = 10⁻³ and −23% for S = 10⁻² over OD 2.0–3.0.
Conclusion. The stray-light mechanism is real but confined to OD > 1.5. It cannot explain any compression at OD 0.3–0.6. The widely repeated warning about the linear range is, for this mechanism, misdirected.
What survives. Two things. First, the high-OD regime matters for anyone fitting late exponential phase or using long path lengths, and the correction is computable from the reader’s stray-light specification. Second, the absence of a stray-light explanation means any observed sub-exponential growth below OD 1 must come from a different mechanism — detector nonlinearity, meniscus and pipetting, evaporation, or a genuine physiological shift — and those are separable by the tests below.
Tests that would settle the residual. Fill-volume series (path length), gain series (detector), and sealed vs open plates (evaporation). If the apparent μ depends on fill volume at fixed gain, the mechanism is optical; if on gain only, electronic; if on sealing, evaporative.
Why this is kept. A hypothesis that fails a quantitative check is more useful than one that is merely plausible. The refinement narrowed a broad claim to a specific, quantified, high-OD regime and removed a mechanism from the list of suspects for the low-OD regime.
R05 · Four-bar conductivity measurements have an unbounded systematic from finite pad conductivity
Refines P09. Source: arXiv:2602.02418.
Claim. In a four-bar geometry the voltage-sensing pads are assumed to draw no current. When the pads have finite conductivity and the film is thin, current spreads into them, and the resulting error is not a contact resistance: it grows without bound as the film and pad conductivities approach each other.
Status of the result. The unboundedness is demonstrated in the cited work; my contribution here is the classification and the test design, not the derivation.
Why it is not a contact resistance. A contact resistance adds a series term that can be measured by varying the pad–film interface. Pad spreading adds a parallel path whose effect grows with the pad’s conductivity relative to the film’s; the two have different scaling with pad area and with film thickness, so a series-resistance calibration cannot remove it.
Discriminating observation. Extracted film conductivity should depend on pad metal at fixed film, with the discrepancy scaling as the conductivity ratio. It should also depend on pad width at fixed metal.
Refutation. Extracted conductivity invariant to pad material across a 100× conductivity range refutes the mechanism. So does invariance to pad width at fixed metal.
Limits. The effect is largest for highly conductive films on highly conductive pads — precisely the regime of transparent conductors and of metal films measured with lithographic leads. For semiconducting films with low-mobility pads the error may be negligible, which is why the practice survived.
Practical use. Report the pad/film conductivity ratio with every four-bar measurement, and use the van der Pauw geometry where the ratio is unfavourable.
R06 · Cryo-EM particle picking is confirmation-biased, and the bias scales with prior strength
Refines B05. Source: arXiv:2507.03951.
Claim. Template or network picking preferentially accepts images that already resemble the template. The selected subset is enriched in template-like views, which sharpens features present in the template and suppresses features absent from it.
Mechanism. Picking is a detection problem with a prior; the posterior over particle locations is conditioned on the template. The reconstruction then inherits a selection function whose shape is the template’s.
Genuine prediction. The bias is a function of prior strength, so a stronger prior (lower threshold, more distinctive template) should produce a larger discrepancy between the picked subset’s density and an independently determined one. Crucially, the effect is invisible in resolution metrics: a self-fulfilling map can be sharp.
Discriminating observation. Template-swap: reconstruct from subsets picked with two templates that differ in a specified region. The density in that region should differ, and the difference should scale with the templates’ difference. A second test: pick on a dataset before and after 3D classification and compare the maps’ densities for features absent from the initial template.
Refutation. Template-swapped reconstructions identical at the reported resolution refute the mechanism. So does a density that is invariant to picking threshold.
Limits. The cited work quantifies the bias; the magnitude for any given dataset depends on the template’s mismatch to the true particle, which is unknown a priori. The prediction is directional and scaling, not a universal number.
Practical use. Report the picking threshold and template; run a template-swap control on at least one region of interest.
R07 · Division by a noisy denominator manufactures a negative correlation that is routinely read as a trade-off
Refines X47. Source: R1_refinement_numbers.py,
section C.
Claim. For a ratio X/Y with X and Y independent, corr(X/Y, Y) = −CV_Y/√(CV_X² + CV_Y² + CV_X²CV_Y²). The correlation is a property of the coefficient of variation, not of the biology or physics.
What is an identity. The formula is an identity for independent log-normal X and Y. It should be presented as a warning, not a discovery.
What is a genuine prediction. The magnitude is large in realistic regimes: −0.71 at CV = 0.1/0.1, −0.94 at 0.1/0.3, −0.32 at 0.3/0.1. The prediction is that measured “trade-offs” between a per-unit quantity and its denominator should weaken when the denominator’s precision improves, in a way fixed by the CV ratio.
Discriminating observation. Improve denominator precision by a known factor and re-measure the correlation. If the correlation is spurious, it should move along the predicted curve; if real, it should be unchanged.
Refutation. A correlation that is invariant to a tenfold improvement in denominator precision refutes the artefact explanation.
Limits. Log-normal marginals are assumed. With heavy-tailed denominators the exact form changes but the sign and order of magnitude do not. The artefact and a real trade-off can coexist; the test is designed to separate them, not to declare all such correlations spurious.
Practical use. Report CVs alongside any ratio–denominator correlation, and prefer log-ratio models.
R08 · Inverting a mean response through a curved calibration biases the mean input
Refines X45. Source: R1_refinement_numbers.py,
section B.
Claim. For a calibration y = ax + bx², the mean response corresponds to an input above the mean input by ≈ (b/a)·σ_x². The error is second order in variability and first order in curvature.
What is an identity. This is Jensen’s inequality applied to the inverse calibration; the leading term is a Taylor expansion. It is not new mathematics.
What is a genuine prediction. The bias is large enough to matter in routine work: for b/a = 0.5 and CV = 0.6 the mean input is biased by 9%; for b/a = 0.2 and CV = 0.3 by 1.3%. The prediction is that two records with the same true mean but different variability will report different means, with a difference fixed by the curvature.
Discriminating observation. Compare high- and low-variability records at matched true mean; the inferred means should differ by the predicted amount. Alternatively, invert each sample before averaging and compare.
Refutation. Inferred means that are variability-independent for a calibration with measurable curvature refute the mechanism.
Limits. The expansion is first order in curvature; for strongly curved calibrations the full inverse must be integrated. The direction of the bias depends on the sign of b, so the test must use a calibration whose curvature sign is known.
Practical use. Average the input, not the response, whenever the calibration is curved — which is almost always.
R09 · Fitting in log space with additive noise biases the faint end upward and flattens slopes
Refines X15. Source: standard error-propagation result.
Claim. With additive Gaussian noise of variance σ², ln y is biased by −σ²/(2y²) — the bias is largest where y is smallest. A log-space fit therefore bends the faint end upward and reduces the fitted slope.
Status. This is a second-order delta-method identity. Its importance is practical: data spanning decades are almost always fitted in log space.
Genuine prediction. The slope bias is predictable from the relative noise at the faint end: a fit spanning a factor of 100 in y with 10% relative noise at the faint end should have its slope reduced by an amount computable from the noise profile. The prediction is quantitative and checkable on any published dataset that reports per-point uncertainties.
Discriminating observation. Refit the same data with a likelihood that matches the noise model (additive Gaussian in linear space) and compare slopes. The difference should match the delta-method prediction.
Refutation. Slopes that agree between log-space and correct-likelihood fits, for data with substantial relative noise at the faint end, refute the mechanism.
Limits. If the noise is genuinely multiplicative, log-space fitting is correct and the hypothesis does not apply. The test therefore requires knowing the noise model, which is often the unexamined assumption.
Practical use. State the noise model; fit in the space where the noise is homoscedastic.
R10 · Normalising to a reference imports that reference's error into every sample
Refines X09. Source: compositional and reference-normalisation practice.
Claim. When measurements are expressed relative to a reference, the reference’s error is common to all normalised values and does not average out. It appears as a correlation between nominally independent samples.
Mechanism. Normalisation is a division by a random variable; the shared denominator induces a common factor whose variance is the reference’s relative variance.
Genuine prediction. The induced correlation is not removed by increasing the number of samples — only by increasing the reference’s precision or by using a reference-free model. The prediction is that inter-sample correlations should be invariant to which sample is chosen as reference if they are real, and should change sign or vanish under reference permutation if they are artefacts.
Discriminating observation. Reference-permutation analysis: recompute the correlation matrix with each sample in turn as the reference. A real correlation is stable; an artefact moves.
Refutation. Correlations invariant to reference permutation, in a dataset where the reference’s own CV is large, refute the mechanism.
Limits. If the reference is measured much more precisely than the samples (a certified standard), the imported error is negligible and the hypothesis does not apply.
Practical use. Permute the reference before reporting any correlation from normalised data.
R11 · Two interface conductances in a TDTR stack are never separable at one modulation frequency
Refines M03. Source: R1_refinement_numbers.py,
section I (thermal quadrupole).
Claim. In a transducer/film/substrate stack with two interface conductances, a single-frequency measurement determines only a weighted combination. The weighting is set by the resistances’ share of the total thermal impedance, and it changes with modulation frequency only because the film’s heat capacity introduces a frequency-dependent element.
Computation. A 1D thermal-quadrupole model (Al transducer 80 nm, film 100 nm, G₁ = 10⁸, G₂ = 5×10⁷ W/m²K) gives S_G2/S_G1 = 2.02 at 100 kHz falling to 0.55 at 30 MHz. The ratio changes by a factor of 3.7 across the band, which is what makes a multi-frequency fit possible at all.
What is an identity. That two series resistances are inseparable at a single frequency is elementary. The quantitative ratio and its frequency dependence are model results.
Genuine prediction. The fitted G should drift with modulation frequency in the direction and by the amount predicted by the sensitivity ratio, and should stabilise only when the fit includes enough of the band to constrain both terms. A corollary prediction: reported G values for the same interface measured at different frequencies should differ systematically, and the difference should be larger for thinner films (where the heat capacity term is weaker).
Discriminating observation. Frequency-series TDTR on a sample with independently characterised transducer adhesion.
Refutation. A fitted G that is frequency-independent across 100 kHz–30 MHz, for a film thin enough that both boundaries contribute, refutes the mechanism.
Limits. The model is 1D and isotropic; radial heat flow and anisotropic films add terms. The specific ratios depend on the assumed conductivities and are illustrative, not universal.
Practical use. Fit the full frequency series with both conductances free; report the sensitivity ratio so readers can see which boundary dominates.
R12 · Metal–dielectric thermal boundary conductance measured on thin films includes the electron–phonon resistance
Refines M38. Source: two-temperature model practice.
Claim. Heat must equilibrate between electrons and phonons in the metal over a finite length. For metal films thinner than that length, the measured interface conductance includes the electron–phonon resistance in series, and the reported G depends on metal thickness.
Mechanism. Two resistances in series: the interface and the electron–phonon channel. The second scales with thickness, so a thickness series separates them without any additional measurement.
Genuine prediction. G_app should increase with metal thickness and saturate at the electron–phonon length, which is tens of nanometres for typical metals. The saturation length is a material property and should be the same for different substrates at fixed metal.
Discriminating observation. Metal-thickness series at fixed substrate and fixed interface preparation.
Refutation. G independent of metal thickness below 20 nm refutes the mechanism for that system.
Limits. The electron–phonon length depends on temperature and on the electron mean free path, so the test must be at fixed temperature. For thick films the effect vanishes, which is why it survived: most TDTR work uses 60–100 nm transducers where the correction is small but not zero.
Practical use. Use two transducer thicknesses, or correct with the known electron–phonon coupling constant.
R13 · Amorphous films thinner than the dominant phonon mean free path have suppressed thermal conductivity
Refines M45. Source: thermal-conductivity suppression literature.
Claim. Surface scattering removes the low-frequency modes that carry most of the heat in amorphous films. For thicknesses below the dominant mean free path the conductivity is suppressed, and the suppression follows a thickness-dependent function.
Genuine prediction. κ should decrease below a threshold thickness and should be lower for rougher surfaces at the same thickness. The threshold is material-specific and is not a universal constant.
Discriminating observation. Thickness series with independently measured surface roughness.
Refutation. κ independent of thickness above 5 nm for a material whose dominant mean free path is known to be 50 nm refutes the mechanism.
Limits. In amorphous materials the mean free path is ill-defined and is itself a fitted quantity, so the prediction is a functional form rather than a number. Interface layers and oxidation can mimic the effect.
R14 · The constant-phase exponent of a porous electrode is a geometry parameter, not a roughness parameter
Refines C35. Source: porous-electrode transmission-line models.
Claim. In a porous electrode with a distributed RC transmission line, the impedance is a power law over the frequency range where the penetration depth lies inside the pore. That power law mimics a constant-phase element with exponent 0.5–0.9, and the exponent is set by pore geometry.
Genuine prediction. The CPE exponent should change with electrode thickness at fixed material, following the transmission-line prediction; a roughness-driven CPE would be thickness-independent.
Discriminating observation. Thickness series of the same electrode material, with the exponent fitted over a fixed frequency band.
Refutation. Exponent independent of thickness, in a material with measurable pore-length polydispersity, refutes the geometric explanation.
Limits. Both mechanisms can operate at once, and the transmission-line model requires an assumption about the pore-size distribution. The test separates the dominant term, not the two contributions exactly.
R15 · The low-frequency tail of an impedance spectrum is drift, not diffusion
Refines C18. Source: drift artefacts in low-frequency EIS.
Claim. Slow drift of the open-circuit potential over the minutes-to-hours of a low-frequency sweep produces a current ramp whose demodulated quadrature component is ∝ 1/ω — indistinguishable, in a single sweep, from a Warburg or CPE element.
Mechanism. The lock-in demodulation of a linear ramp over an integration time T produces a quadrature signal proportional to the ramp rate times T. Because T is itself proportional to 1/ω for a swept measurement, the artefact is ∝ 1/ω.
Genuine prediction. Reversing the sweep direction should change the sign of the low-frequency tail while leaving the mid-frequency arc intact. A true diffusion element is direction-independent.
Discriminating observation. Forward and reverse sweeps at fixed DC bias, with the DC drift monitored independently.
Refutation. Identical forward and reverse low-frequency tails refute the drift explanation.
Limits. A genuine Warburg and drift can coexist; the direction test separates them only if the drift is monotonic over the sweep time. Galvanostatic operation changes the artefact’s form.
R16 · Open-circuit voltage cannot give state of charge until the diffusion relaxation is complete
Refines E28. Source: OCV–SOC hysteresis literature.
Claim. OCV is read after a fixed rest and mapped to SOC. Diffusion relaxation in thick electrodes takes hours; a one-hour rest leaves a voltage offset whose sign depends on the preceding current direction.
Mechanism. The electrode’s solid-phase concentration profile relaxes with a spectrum of time constants set by particle radius and diffusivity. The terminal voltage is a functional of the surface concentration, which is not the mean concentration until relaxation completes.
Genuine prediction. The OCV–SOC curve measured after charge lies systematically above the curve measured after discharge, and the gap closes with rest time in a way that scales as the particle diffusion time. The prediction is quantitative for a given particle size: the gap should shrink by a factor of e when rest time increases by the diffusion time.
Discriminating observation. Rest-duration series after charge and after discharge on the same cell, at several SOCs.
Refutation. OCV independent of prior current direction after a one-hour rest refutes the mechanism for that electrode.
Limits. For thin electrodes or high-diffusivity materials the relaxation is fast and the artefact is negligible. The mapping from rest time to relaxation is model-dependent.
Practical use. Report rest time and prior direction with every OCV–SOC point; use a relaxation model rather than a lookup table.
R17 · Surface-plasmon resonance affinities are limited by mass transport, and the limitation survives a good global fit
Refines B16. Source: mass-transport limitation in SPR.
Claim. Mass-transport limitation produces a response that is well fitted by a 1:1 model with a reduced k_on, because both effects saturate. The fit residual is small while k_on is underestimated, by up to an order of magnitude.
Mechanism. The analyte flux to the surface is bounded by diffusion through the depletion layer. When the reaction is faster than transport, the observed rate is the transport rate, and the 1:1 model absorbs it.
Genuine prediction. k_on should increase with flow rate and saturate, and should approach the true value when ligand density is reduced. The dependence on ligand density is the discriminating one: transport limitation scales with the Damköhler number, which is proportional to ligand density.
Discriminating observation. Flow-rate and ligand-density series on the same interaction.
Refutation. k_on invariant to flow rate and ligand density refutes the mechanism.
Limits. At very low ligand density the signal-to-noise falls, so the test has a practical floor. The Damköhler number must be estimated from the measured k_on, which is circular unless an independent estimate of the transport coefficient is available.
Practical use. Report ligand density and flow rate; run one injection at a different flow rate as a control.
R18 · Bolometer time constants measured in situ are biased by the optical load they are meant to characterise
Refines P22. Source: electrothermal TES modelling.
Claim. Transition-edge sensor response is electrothermal; the effective time constant depends on bias power and hence on optical load. A constant measured at high calibration load is shorter than the sky-load value, biasing beam-map deconvolution for fast scans.
Genuine prediction. Time constants should vary with controlled load (aperture cover) with a slope predictable from the electrothermal model, and map transfer functions should depend on scan speed accordingly.
Discriminating observation. Load-variation campaign, with the same detector measured under several aperture states and at several scan speeds.
Refutation. Time constant invariant to load within 1% refutes the mechanism for that detector.
Limits. The electrothermal model requires the detector’s bias circuit parameters, which are not always reported. The prediction is directional and scaling, not a universal constant.
R19 · NV thermometry has a strain–temperature degeneracy whose size depends on the crystal cut
Refines P08. Source: Zenodo 21299294 (pre-registered NV thermometry comparison).
Claim. Axial strain and temperature both shift the zero-field splitting D. The two contributions are collinear for a [100]-cut ensemble but not for [111], so a single-axis measurement reports a temperature error proportional to the local strain.
Mechanism. The spin-strain tensor’s axial component adds to D; the temperature coefficient is also a shift in D. For one orientation the two enter with the same sign and cannot be separated; for another they differ in sign or magnitude.
Genuine prediction. Temperature readings from [100] and [111] ensembles on the same sample should disagree under applied stress and agree stress-free. The disagreement should be linear in the applied strain with a coefficient set by the spin-strain coupling.
Discriminating observation. Dual-orientation diamond with a piezoelectric strain stage.
Refutation. Both cuts tracking the same temperature within 0.1 K under 10⁻⁴ strain refutes the mechanism.
Limits. The comparison requires two ensembles in the same local strain field, which is itself difficult. The published comparison of 23 methods reports geometric blind spots that are consistent with this mechanism but do not isolate it.
R20 · A forgotten trend deflates error bars, but only modestly
Refines X13. Source: R1_refinement_numbers.py,
section G.
Claim. Error bars on a time average use an autocorrelation time estimated from the same record. Detrending removes the variance that dominates the uncertainty.
Calculation. For a record with a linear ramp of amplitude a and noise σ, the true standard error of the mean is √(a²/12 + σ²), while the naive estimate from residuals is σ. The underestimate factor is 1.01 for a/σ = 0.5, 1.15 for a/σ = 2, and 1.53 for a/σ = 4.
Correction to the original claim. The effect is smaller than the original entry implied. It matters only when the trend dominates the noise (a/σ ≳ 2) and even then deflates the error bar by less than a factor of two. The original entry overstated the generality.
What survives. For records where the trend is several times the noise — which is exactly the case in long environmental and astrophysical series — a 50% underestimate is worth correcting, and the correction is analytically available.
Discriminating observation. Compare single-record error estimates against ensemble estimates from replicate records; the ratio should follow the predicted curve in a/σ.
Refutation. Ratio equal to one for a/σ = 4 refutes the mechanism.
Limits. The calculation assumes a linear trend and white noise. Red noise changes the arithmetic but not the direction.
R21 · Inverse-variance weighting with estimated variances does not bias the mean, but understates its uncertainty
Refines X33. Source: R1_refinement_numbers.py,
section H (Monte Carlo, 40 000 trials).
Original claim. Weighting by inverse estimated variance biases the weighted mean toward points whose noise fluctuated low.
What the simulation shows. The mean is unbiased (mean −0.000 to 0.005 across 2–30 degrees of freedom). What changes is the spread: the estimator’s standard deviation is 0.686 at 2 dof, 0.582 at 5, 0.539 at 10, 0.511 at 30, against a nominal 0.5. The quoted uncertainty is therefore 73%, 86%, 93%, 98% of the true spread.
Correction. The hypothesis as stated is wrong about the mean and right about the uncertainty, with a magnitude that depends on the degrees of freedom used to estimate each variance.
Genuine prediction. Reported weighted means with per-point variances estimated from few replicates should show excess scatter across repeated experiments, by the predicted factor, while their central values remain unbiased. The prediction is testable on any replicated meta-analysis.
Discriminating observation. Replication scatter versus quoted uncertainty, stratified by the dof used per point.
Refutation. Scatter matching the quoted uncertainty for dof = 2 refutes the mechanism.
Limits. The simulation uses Gaussian data and four measurements per dataset; other configurations shift the numbers but not the structure.
R22 · Censoring at a detection limit biases trends in a direction set by the trend
Refines B39 / X11. Source: censored-data methods.
Claim. Substituting a constant for values below the limit compresses the low end. If the quantity is increasing over time, the censored fraction falls, so the substituted values occupy less of the record and the trend is biased by an amount that depends on the censoring fraction’s trend.
Mechanism. Substitution is a biased estimator of the conditional mean below the limit; the bias enters the trend through the changing mixture weight.
Genuine prediction. The trend estimated with a censored-data likelihood should differ from the substitution estimate by an amount that scales with the change in censoring fraction across the record, not with its average. A record with constant censoring shows no bias even if heavily censored.
Discriminating observation. Reanalysis of time series with reported detection limits and censoring fractions.
Refutation. Trends unchanged under a censored likelihood, in a record whose censoring fraction changes by a factor of two, refutes the mechanism.
Limits. The bias direction depends on the substitution rule (LOD/2, LOD/√2, deletion); the prediction should be made for the rule actually used, which is often not stated.
R23 · Low-cost particle sensors report a humidity-driven mass that mimics winter pollution
Refines E01. Source: low-cost sensor correction literature.
Claim. Hygroscopic growth increases the optical diameter at high relative humidity. The reported PM₂.₅ rises with RH at constant mass, and a co-location calibration over a narrow humidity range cannot identify the coefficient.
Mechanism. Optical scattering scales steeply with diameter; deliquesced aerosol scatters more per unit dry mass. The growth factor is a known function of RH (κ-Köhler), so the artefact is predictable from the aerosol’s hygroscopicity.
Genuine prediction. The sensor/reference ratio at fixed reference PM should correlate with RH with a slope independent of site, and the slope should scale with the aerosol’s κ. The prediction is that a network’s seasonal cycle contains an RH-shaped component that is absent from the reference.
Discriminating observation. Co-location data spanning a full humidity range, ideally at two sites with different aerosol hygroscopicity.
Refutation. Ratio independent of RH across sites refutes the mechanism.
Limits. The growth factor depends on composition, which varies; a single correction will not be exact. The prediction is a systematic shape, not a point correction.
R24 · DSC onset temperatures shift with scan rate by an amount set by the pan–sample time constant
Refines C25. Source: R1_refinement_numbers.py,
section F.
Claim. The measured onset is the true onset plus τβ, where τ is the effective pan–sample thermal time constant and β the scan rate. With τ = 5–20 s the shift is 0.4–6.7 K over 5–20 K/min.
Mechanism. The sample’s temperature lags the programmed temperature by τ; the onset criterion is applied to the programmed value.
Genuine prediction. The onset should be linear in scan rate with slope τ, and τ should scale with sample mass and with pan–sample contact. Zero-rate extrapolation recovers the true onset only if the relation is linear, which fails when the pan’s thermal resistance is comparable to the sample’s.
Discriminating observation. Scan-rate and mass series on the same material.
Refutation. Onset independent of scan rate for a sample with a measurable thermal lag refutes the mechanism.
Limits. The time constant is not directly measurable in a standard DSC; the test infers it from the slope, which is the quantity of interest.
R25 · Permeability assays measure the unstirred water layer as much as the monolayer
Refines B42. Source: R1_refinement_numbers.py,
section E.
Claim. P_app⁻¹ = P_true⁻¹ + P_UWL⁻¹ with P_UWL = D/h. For a 1000 µm unstirred layer and D = 5×10⁻⁶ cm²/s, P_UWL ≈ 5×10⁻⁵ cm/s; a compound with P_true = 5×10⁻⁴ cm/s is reported at 9% of its true value.
What is an identity. The series-resistance formula is elementary.
What is a genuine prediction. The distortion is compound-dependent and predictable: the ratio P_app/P_true falls from 0.91 for P_true = 5×10⁻⁶ cm/s to 0.09 for 5×10⁻⁴ cm/s at h = 1000 µm. Shaking reduces h and therefore moves the ratio toward 1 for permeable compounds only. The prediction is that in-vitro/in-vivo correlations improve when P_app is corrected for the measured h, and that the correction matters only above a threshold permeability.
Discriminating observation. Shaking-rate series across a compound panel of graded permeability.
Refutation. P_app independent of shaking rate for permeable compounds refutes the mechanism.
Limits. D must be known or estimated; h is not directly measured in most assays. The series model assumes a uniform layer.
Practical use. Report shaking rate; include at least one high-permeability reference compound to expose the layer.
R26 · Dielectric breakdown strength scales with electrode area, so reported strengths are not material properties
Refines M33. Source: R1_refinement_numbers.py,
section D.
Claim. Breakdown is defect-initiated, so the Weibull scale parameter scales as (A/A₀)^(−1/m). For m = 10, a 100× area increase reduces the reported strength to 63% of the small-area value; for m = 5, to 40%.
What is an identity. The area scaling is a standard result of weakest-link statistics.
What is a genuine prediction. The slope 1/m is measurable from an area series and is a property of the defect population, not of the material’s intrinsic strength. The prediction is that interlaboratory scatter in reported breakdown strength should collapse when normalised to a common area, and that the collapse should be exact only when the defect populations are similar.
Discriminating observation. Area series in one laboratory, then comparison of normalised values across laboratories.
Refutation. Scale parameter independent of area, for a material with a measurable defect density, refutes the mechanism.
Limits. The Weibull model assumes a single defect population; multimodal defect distributions produce curvature in the area plot that the single-m gives incorrectly.
Practical use. Report electrode area and the fitted Weibull modulus with every breakdown strength.
R27 · Interferometric total flux is set by the shortest baselines, so flux errors are not constrained by closure quantities
Refines P42. Source: wide-field interferometric calibration.
Claim. Total flux is carried by the shortest baselines, where the primary beam and the wide-field term are largest. A fractional error in the shortest-baseline gain moves total flux without changing image shape, and closure quantities do not constrain it.
Mechanism. Closure phase and closure amplitude are invariant to station-based gains. A common gain error across all short baselines is therefore invisible to closure but changes the total flux.
Genuine prediction. Total flux should correlate with the shortest-baseline gain solution, and should differ between imaging with and without short baselines (tapering). The prediction is that the flux scale error is a function of the array’s minimum baseline, not of its maximum.
Discriminating observation. Re-image the same dataset with and without tapering, and compare total flux against an independent (single-dish) measurement.
Refutation. Total flux invariant to tapering and to shortest-baseline gain refutes the mechanism.
Limits. For resolved sources, tapering also changes the recovered flux through missing extended structure, so the two effects must be separated with a source model.
R28 · Cycloheximide chase half-lives are shorter than degron-based half-lives for the same protein
Refines B13. Source: comparison of translation-inhibition methods.
Claim. Cycloheximide perturbs the cell within minutes; the chase’s t = 0 is defined after the perturbation, so the fitted decay includes a fast stress component.
Genuine prediction. Half-lives from cycloheximide should be systematically shorter than from an auxin-inducible degron for the same protein, and the difference should scale with the induction of a stress marker. The prediction is comparative: the two methods should converge for proteins whose degradation is stress-insensitive.
Discriminating observation. Side-by-side chase with both methods on a panel of proteins of known stability, with a stress-marker readout.
Refutation. Half-lives agreeing across methods for stress-sensitive proteins refutes the mechanism.
Limits. Degron systems have their own artefacts (auxin effects, tag effects on stability). The test is a triangulation, not a gold standard.
R29 · Ozonesonde records contain an instrument-generation trend from pump efficiency
Refines E15. Source: ozonesonde homogenisation literature.
Claim. Pump efficiency falls with reduced pressure and with contamination. The correction depends on sonde age and on the launch profile, so a fleet that changes generation acquires a trend confounded with the atmospheric one.
Genuine prediction. The sonde-to-satellite or sonde-to-lidar ratio should differ between generations after the standard correction, and the difference should be largest in the upper troposphere–lower stratosphere where the pump works hardest.
Discriminating observation. Intercomparison campaigns with sondes from two generations launched together.
Refutation. Ratios independent of generation after standard correction refute the mechanism.
Limits. Generation changes usually coincide with other changes (sensor chemistry, launch practice), so isolation requires a deliberate side-by-side campaign.
R30 · A global fit absorbs unmodelled physics into whichever parameter is most flexible
Refines X30. Source: model-dependence of fitted constants.
Claim. The covariance matrix reflects only the parameters in the model. Unmodelled physics is absorbed by the most flexible parameter, which then appears both precise and shifted. Its precision is a statement about the model’s rigidity, not about the measurement.
Mechanism. The fit projects the model discrepancy onto the parameter space along the direction of least resistance. That direction is set by the model’s structure, not by the data.
Genuine prediction. The parameter should shift by more than its quoted uncertainty when the model is extended with physically motivated terms, and the shift should be largest for the parameter with the largest flexibility (as measured by the width of its posterior under the original model). A corollary: parameters with narrow posteriors are the most likely to be absorbing discrepancy, not the best measured.
Discriminating observation. Model-extension sensitivity on a published fit: add one physically motivated term and record which parameter moves.
Refutation. Parameters stable under physically motivated extensions refute the mechanism.
Limits. The test requires a physically motivated extension, which is exactly what may be missing. A parameter that is stable under every extension the community has tried is evidence for robustness, not proof of it.