Skip to content
Science Journaling Club

Volume 2, Issue 3March 2026

Spring 2026

Two reviews asking what the evidence actually supports rather than what the headline said, plus a dinosaur decline that turns out to track exposed rock rather than dinosaurs, eighty-one cubic kilometres of groundwater nobody had counted, and a mantle plume whose rhythm is set by the plates above it.

This issueSJC 2.3
Published
March 2026
Articles
8
Research
3
Field notes
5
With code
5

Contents

8 articles, in the order the club meant them to be read.

  1. ResearchMethods

    A Greenhouse Model You Can Check by Hand, and Where It Stops Being Honest

    The layer model is how almost everybody first meets the greenhouse effect, and it is a good model. We asked it a question it is rarely asked: how many layers does Earth need? The instrument was a Python file the club wrote and ran on a laptop, and nothing was measured by us at any point, because here the computation is the experiment. Solving the NN-layer grey radiative-equilibrium system as a linear system for every NN from 0 to 10 reproduced the exact analytic answer Ts=Te(N+1)1/4T_s = T_e (N+1)^{1/4} to a largest discrepancy of 5.7 × 10−14 K, a relative error of 1.22 × 10−16, which is 0.55 of one double-precision epsilon. The zero-layer case returned 255.212 K against the textbook's 255 K. The answer to our question is that Earth needs 0.622 of a layer, which is not a number of layers at all: one perfectly absorbing layer gives 303.500 K, overshooting the real surface by 15.50 K, or 47% of the entire 32.8 K greenhouse effect the model exists to explain. Replacing the perfect absorber with a tunable emissivity fixes this and yields ε∗=0.7667±0.0122\varepsilon^\ast = \mathbf{0.7667 \pm 0.0122} from 200,000 Monte Carlo trials over published uncertainties in the solar constant, the albedo and the surface temperature. Then we measured what grey costs. Against a two-band model that separates the 8 to 12 µm window from the rest of the spectrum, the grey assumption overstates the warming from a doubling of absorber by a factor of 2.02 ± 0.01, and hides a saturation ceiling: a hundredfold increase in band opacity is worth 2.64 K and nothing more. Worse, a single-layer two-band model can reproduce the observed surface temperature or the observed 22 W/m² of surface emission that escapes directly through the window, and not both. Forced to match the escape, it runs 12.0 K too hot. Seed 20260321.

    Spring 2026Read 35 minSections 12

  2. ResearchOriginal Research

    Computing the Sunlight That Started and Ended the Ice Ages

    The amount of sunlight falling on the top of the atmosphere at any latitude on any day is a geometry problem with four inputs and no physics in it at all. We solved it, and then ran it across the past million years. The instrument was a Python file the club wrote; the only thing we did not compute ourselves is the orbit, which comes from the published La2004 solution of Laskar and colleagues, downloaded on 14 September 2026 and cached with its checksum. Summer solstice insolation at 65°N came out at 477.94 W/m² today and has ranged over 128.25 W/m² in the past million years, from 430.53 at 231 kyr BP to 558.78 at 958 kyr BP, a swing of 26.0% of the mean. Before trusting any of that we set the formula against six closed forms derived independently: the global annual mean came out 340.297512 against an exact 340.297470 W/m², the equatorial annual mean agreed with an elliptic-integral expression to 5.7×10−14, and a Monte Carlo integration over the sphere that never touches the daily formula landed 1.50 standard errors away on 4,000,000 samples. Transcribing Laskar's own Fortran subroutine and running it beside ours gave a maximum disagreement of 2.8×10−13 W/m² over a grid of 130,501 points. Varying each orbital parameter alone across its million-year range, the perihelion angle moves 65°N solstice insolation by 114.53 W/m² at maximum eccentricity and by nothing at all at zero eccentricity, obliquity by 37.97, and eccentricity on its own by 46.98. The spectrum of the resulting curve puts 81.26% of its variance in the precession band and 17.74% in the obliquity band, and it recovers the precession triplet at 23.83, 22.24 and 18.89 kyr and the obliquity line at 41.71 kyr, with 0–5 Myr windows sharpening those to 23.701 and 40.992 kyr. The 75 to 135 kyr band holds 0.07%. The nine glacial terminations of the LR04 age model sit a mean 1.78 kyr from the nearest maximum of our curve against 5.52 kyr for random dates, which is an alignment no set of nine random dates matched in 20,000 draws. It is also partly circular, and the section on the strongest objection says why. The curve offers 46 maxima over the million years. The ice offered nine. Seed 20260320.

    Spring 2026Read 32 minSections 12

  3. ResearchOriginal Research

    Why the Planet Has Not Finished Warming Yet

    Heat added at the top of the atmosphere does not warm the surface immediately, because most of it goes into the ocean first. We measured how long the delay is and how much warming it hides. The instrument was a two-box energy balance model the club wrote and integrated on a laptop; no measurement of the physical world was made by us at any point, and the computation is the whole experiment. With the CMIP5 multimodel mean parameters of Geoffroy et al. [2] the model has two response times, 3.9486 yr and 247.75 yr, a ratio of 62.7. Those numbers are not fitted. They fall out of the eigenvalues, and we recovered them from the model's own impulse response by log-linear fitting to 1.4×10−13 yr. Runge-Kutta integration, eigendecomposition and the closed form agree to 1.6×10−14 K across 61 time points spanning almost five decades of time; setting the heat uptake to zero collapses the system to a single box with the exact timescale C/λ = 6.4602 yr, to 1.8×10−13 K; and the integral of the net imbalance matches the heat stored in the two boxes at every one of 600,000 steps, worst residual 2.5×10−11 W yr m−2. Equilibrium sensitivity is 3.05 K and transient response 1.90 K, a ratio of 0.622, against IPCC AR6 best estimates of 3.0 K and 1.8 K [14]. One hundred years after an abrupt doubling the model has delivered 73.61% of its eventual warming and still owes 0.806 K, essentially all of it carried by the slow mode: the fast mode's share of the remainder is 2×10−9%. Sweeping the heat uptake efficiency γ and the feedback λ across their published ranges moves the realized fraction from 60.8% to 87.4%. A Monte Carlo over the intermodel spread, 393,758 retained trials, gives a committed warming at 100 years of 0.957 ± 0.001 K mean and 0.772 K median. The model fails in one place and we report it: driven by an idealised historical forcing tuned to the observed energy imbalance, it produces 1.41 K of warming by 2019 where the observational assessment gives 1.09 K, an overshoot of 29%. That is the pattern effect, and it is a fault in the constant-λ assumption rather than in the lag.

    Spring 2026Read 31 minSections 12

  4. Field noteMeta-analysis

    Not Every Shell Dissolves: Sorting the Ocean Acidification Response Literature by Who Is Actually Affected

    We extracted 23 calcification and growth effect sizes from 16 published ocean acidification experiments, pooled them with a DerSimonian-Laird random-effects model, and asked whether taxonomic group explains why those experiments disagree with each other as violently as they do. The pooled response is a 13.0% reduction in calcification or growth (95% CI −31.0% to +9.6%), which is the number most readers arrive wanting and the number the rest of this article spends its length qualifying. The interval contains zero. Heterogeneity is extreme: I² = 98.8%, τ² = 0.278, Q = 1860 on 22 degrees of freedom, and the 95% prediction interval for the next experiment anybody runs stretches from −72% to +167%. Taxon is the obvious candidate, and the naive fixed-effect subgroup decomposition appears to confirm it in spectacular fashion, returning Qbetween = 754 on 7 degrees of freedom. We then fed that same test simulated data in which every taxon had been assigned identical true effects, and at our observed level of within-taxon disagreement it announced a significant taxon difference in 97.9% of runs where none existed. The correctly specified mixed-effects moderator test gives QM = 3.14 on 7 degrees of freedom, p = 0.87, so taxon explains essentially none of the heterogeneity, and experiments on one organism disagree with each other about as hard as experiments on separate phyla do. Eleven students extracting numbers from archived data files are not a systematic review with two independent screeners, and we say plainly below what that costs the estimate and what it does not. So the result worth reporting is the variance itself, together with its flat refusal to sort along the taxonomic lines that everyone, ourselves included, expected of it.

    Marine BiologyRead 42 minSections 13

  5. Field noteSystematic Review

    Can a Cup of Pond Water Tell You What Lives in It? Reviewing Environmental DNA Detection Rates

    Filter some water, lyse what the membrane caught, run a PCR, and you are supposed to learn which animals were recently swimming there. We went looking for every study we could open that surveyed the same water bodies twice, once with a filter and once with a net or a torch or a camera, and that reported a detection probability for both. We found twelve we could pool and seven more we could not. On a random effects model the odds that one eDNA survey unit detects an occupied site are 2.05 times the odds for one conventional survey unit, with a 95% interval of 0.95 to 4.41, which does not exclude a tie. The heterogeneity is the actual result: I² = 80.5%, and the interval predicting the next study runs from an odds ratio of 0.15 to 27. Six of the twelve contrasts favour the filter and six favour the net. The strongest single pattern is taxonomic: amphibians sit at an odds ratio near four, reptiles below one, and no protocol tuning we can see closes that gap. Then we set the pooling aside and look at the arithmetic nobody can get around: with a per-sample false positive rate of 0.02 and a true occupancy of 0.02, a single positive is real only 38% of the time. Requiring two positives out of two lifts that to 0.95 and drops sensitivity from 0.60 to 0.36, and that trade is the method rather than a small print item.

    Conservation MethodsRead 38 minSections 13

  6. Field notePaper Analysis

    The Dinosaurs Were Fine. The Rocks Weren't.

    Every raw count of dinosaur fossils in North America shows the same shape: diversity climbs through the Campanian, peaks around 75 million years ago, and falls for the final nine million years before the asteroid. That fall has anchored the "dinosaurs were already doomed" story for decades. In April 2025, a team led by Christopher Dean published a paper in Current Biology arguing that the fall is mostly not about dinosaurs at all — it is about rock [1]. Borrowing a technique built for bird and amphibian surveys, the authors fit a Bayesian occupancy model that separates two things a raw fossil count always confuses: whether a dinosaur clade was actually present at a site, and whether the sediment survives today in a form that could tell you so. Modelled occupancy across four major clades holds roughly flat through the interval; what declines, tracking the shrinking area of exposed, datable outcrop and modern land cover, is the probability of detecting a fossil at all [1]. This note walks through how paleontologists count something that has been dead for 66 million years, what a "sampling artifact" means precisely enough to test for, and why the occupancy-versus-detection split is a sharper tool than the correction techniques the field has tried since 1984. We also do our own arithmetic on the numbers the paper and its coverage report — the club's own arithmetic, not the authors' — and it turns up something almost too tidy: the reported peak sits exactly at the midpoint of the study's 18-million-year window.

    PaleobiologyRead 24 minSections 13

  7. Field notePaper Analysis

    Eighty-One Cubic Kilometers of Water the Cascades Never Reported

    Nobody drilled a single well for this discovery. A team led by Leif Karlstrom compiled temperature readings from 72 pre-existing geothermal drill holes punched into the Oregon Cascades decades ago for entirely different reasons, and looked for the sections where temperature stopped rising with depth the way rock is supposed to warm you as you go down. A gradient under 20 °C per kilometer, or one that runs flat or backward, means something is hauling cold surface water downward fast enough to outrun the rock's own heat. Krige those anomalies across the range, multiply by a measured porosity of 6%, and the sum is at least 81 km³ of standing groundwater under the volcanic crest — a number the authors are explicit is a floor, not a ceiling, since their drill holes never mapped the "rain curtain" edge to edge and their arithmetic makes no allowance for water flowing back up [1]. We rebuild the physics that makes a thermometer see water a well cannot, work through why the same aquifer disappears from ordinary stream-gauge bookkeeping (a separate, independent estimate from regional water balance puts the flux at roughly 11 km³ a year, eleven times what the mapped springs alone account for [1]), and then do our own arithmetic on the number 81. It runs to about 2.3 Lake Meads at that reservoir's full-pool capacity, our conversion of a public Bureau of Reclamation figure, not the paper's own comparison [1,12]. At the springs' own discharge rate that volume would take roughly 81 years to fully turn over; at the water-balance rate, about 7 years. Both numbers describe water a region has been managing without, for however long people have been managing it at all.

    HydrogeologyRead 23 minSections 11

  8. Field notePaper Analysis

    A Pulsing Mantle Plume, Timed by the Plate Above It

    Under the Afar triple junction — the one spot on the planet where three tectonic plates are actively pulling apart from a single hinge — a mantle plume is supposed to rise the way textbook plumes rise: one hot column, one radial signature fading smoothly outward. Watts and colleagues measured isotope and trace-element ratios in more than 130 young volcanic rocks spanning all three rift arms and sorted them, with 93 previously published analyses, into groups by k-means clustering on 13 geochemical and geophysical variables [1]. No bullseye appeared. Instead the same handful of chemical "cluster" identities repeat along each rift arm at a fixed spacing, and that spacing is not the same everywhere: about 50–100 km in the Main Ethiopian Rift, versus 150–200 km in the Red Sea Rift, a rift spreading two to four times faster over notably thinner crust. We rebuild that comparison, and then do our own arithmetic — not the paper's — converting each rift's spacing back into an implied time between pulses using its own reported spreading rate. Both arms land in the same window, roughly 10 to 19 million years, despite differing four-fold in kilometres. The paper's own reading is that the plate is not a passenger riding on top of the plume; it is setting the plume's rhythm — a reversal of which direction the causal arrow usually points in rift-plume stories.

    GeodynamicsRead 19 minSections 13