Everything we have published

All articles

Field notes explain somebody else's recent paper for a reader coming to it cold, and usually come with a model you can run in the browser. Research articles report work the club did itself and belong to a quarterly issue. Use the filters to narrow the list.

ResearchMethods
How to Lose a Planet by Choosing the Wrong Integration Method
Simulating gravity is one equation applied over and over: acceleration equals minus $GM$ times the position vector, divided by the cube of the distance. We integrated a single two-body Kepler orbit (eccentricity 0.6, one hundred steps per orbit, a resolution that looks perfectly reasonable) with three standard stepping rules, forward Euler, velocity Verlet and fourth-order Runge-Kutta (RK4), and checked every trajectory against the exact solution of Kepler's equation rather than against each other. At that resolution, Euler's energy flips sign within the first orbit (from −0.500 to +0.083) and never recovers. Velocity Verlet's energy error oscillates in a bounded band between roughly 0% and 2.99% for the full 300 orbits we ran, with no measurable trend (slope consistent with zero at 0.0 of its own standard error), exactly as symplectic theory predicts, and its angular momentum is conserved to $2.4\times10^{-14}$ relative error, 107 machine epsilons, indistinguishable from pure floating-point roundoff. RK4 drifts too, just far more slowly and in the opposite direction: it crosses 1% energy loss only after 47.6 orbits and 5% after 234.2 orbits, a real secular trend (21,915 standard errors from zero) that quietly contracts the orbit rather than expanding it. Fitting the measured convergence order across a range of timesteps gives 0.952 for Euler, 1.998 for Verlet and 4.166 for RK4 against textbook values of 1, 2 and 4; repeating the fit across 40 Monte Carlo trials at random eccentricities between 0.10 and 0.60 gives 0.987 ± 0.002, 2.0015 ± 0.0002 and 4.118 ± 0.005, each a small, real, explainable disagreement with the textbook exponent that we report rather than round away. Giving Euler four times the timestep resolution of RK4, so both cost the same number of force evaluations per orbit, does not rescue it: at eccentricity 0.6 it still loses 1% of the orbital energy within about a thousandth of an orbit, a failure of stability, not merely of accuracy, that a bigger budget cannot buy back. We also flag one place our own title oversells the result: at this configuration Euler's planet does not fall into the sun. It gains energy and flies apart, the better-documented and opposite failure mode; the "spiral into the sun" image turns out to describe RK4's slow contraction far better than it describes Euler's fast escape. Master seed 20260901. Total run time 65.0 seconds.
Fall 2026Read 24 minSections 12
ResearchOriginal Research
Five Places to Park a Spacecraft, Two of Which Are Stable for a Strange Reason
Two bodies orbit their common centre of mass in a circle. In the frame that rotates with them there are exactly five points where a third, massless body can sit still: three on the line through the two masses (L1, L2, L3), and two forming equilateral triangles with them (L4, L5). We located all five numerically for five mass ratios μ spanning five orders of magnitude, from Sun-Earth (μ = 3.00×10⁻⁶) to an illustrative equal-mass pair (μ = 0.5), by two independent numerical methods that agree with each other to machine precision (2.2×10⁻¹⁶) and with the classical small-μ series for the collinear points to four or more significant figures, and by exact algebra for the triangular ones (residual <4×10⁻¹⁶). Eigenvalues of the linearised equations of motion show L1, L2 and L3 unstable at every mass ratio tried, and show L4/L5 stable only below a critical mass ratio that our own numerical bisection places at μ_crit = 0.03852085, against the classical Gascheau (1843) / Routh (1875) value of 0.03852090, a difference of 4.4×10⁻⁸ consistent with the square-root ill-conditioning of an eigenvalue problem at a genuine matrix degeneracy, not a bug. Direct integration of test particles (fixed fourth-order Runge-Kutta, Jacobi constant conserved to <4×10⁻¹³ relative up to first escape in every run) shows the same particle, same perturbation, same code, staying within 1000 times its starting displacement for the whole run at Earth-Moon's L4 and leaving it by t=15.0 at Pluto-Charon's L4, whose mass ratio (0.1087, 2.82× the critical value) exceeds threshold. A Monte Carlo experiment of 1,500 randomly directed perturbations per mass ratio, seeded at 20260914, put the escape probability at L4 at 0% ± 0% at 0.95×μ_crit, 18.8% ± 1.0% at 0.98×μ_crit, and 100% ± 0% by 1.20×μ_crit, with the running estimate visibly converged by a few hundred trials. The stability of L4/L5 has nothing to do with sitting in a potential well: they sit on a hill of the effective potential at every mass ratio, stable or not, and what changes at μ_crit is purely how the Coriolis force, which does no work and appears in no potential, organises the motion around that hill.
Fall 2026Read 31 minSections 12
ResearchOriginal Research
Measuring Exactly How Unpredictable Three Gravitating Bodies Are
Three point masses under Newtonian gravity are the textbook example of chaos, and the claim is usually left at that: asserted, not measured. We built a 4th-order symplectic integrator (Yoshida's composition of velocity-Verlet leapfrog steps [7]), validated it against the analytic two-body Kepler orbit to 5.1×10⁻¹² relative error and against the published figure-eight periodic solution [3][4] to a residual of 1.4×10⁻⁴ after one full period, then measured the largest Lyapunov exponent of three configurations by Benettin's two-trajectory renormalisation method [11]. The exact figure-eight choreography is almost not chaotic at all: λ = 0.0374 ± 0.0014, a Lyapunov time of nearly 27 time units, over four full periods. The 1913 Pythagorean three-body problem [5][6], famous for a violent triple encounter, gives λ = 0.472 ± 0.007, a Lyapunov time under three time units, but only once we replaced a fixed integration step with one tied to the local free-fall time: the fixed step we tried first reported an energy that had grown by a factor of 350, a failure this article reports rather than hides. Forty randomly drawn bound triangles gave a heavy-tailed spread of exponents, median 1.326, mean pulled up to 2.545 by rare close encounters, translating to a prediction horizon of about 10.4 time units before a one-part-in-a-million initial uncertainty grows to the size of the system. Across 1,020 Monte Carlo trials at six initial energies, the fraction of triples that ejected a body within 20 time units rose from roughly 15–26% for bound systems to 62% near the escape threshold and 81% for unbound ones, with standard errors from the trial counts and a convergence plot showing the estimate settling as trials accumulate. A step-size sensitivity check moved the headline ejection fraction from 0.22 to as high as 0.34 under a four-times-coarser step, the single largest source of uncertainty in this study, reported plainly rather than tuned away. Seed 20260914.
Fall 2026Read 26 minSections 12
Field noteExplainer
Neptune's Missing Aurora Was Hiding in the Tropics
Jupiter, Saturn and Uranus have all been watched for decades in the infrared light of a small molecular ion called H₃⁺, which glows wherever a giant planet's magnetic field funnels charged particles into its upper atmosphere. Neptune refused to show it. In March 2025 a team led by Henrik Melin reported that JWST had finally caught the ion, and caught with it the first clear infrared aurora at Neptune, sitting between 30 and 60 degrees south rather than over a pole [1]. The same spectrum reads the upper atmosphere at 358 ± 8 K. Voyager 2 measured 750 ± 150 K there in 1989, roughly twice as hot. We rebuild the geometry ourselves. A tilted dipole traced by hand puts the auroral oval at 56°S to 30°S, which is 86% of the observed band and misses its centre by 2.6 degrees, and a Boltzmann calculation of the ion's emission says the glow at 358 K is 0.51% of what it was at 750 K, a factor of 198. Sections 4 through 7 hold the club's own simplified model rather than the paper's analysis. Section 9 says where it breaks.
Planetary AstronomyRead 33 minSections 12
Field notePaper Analysis
Forty-Five Editors, Thirty Percent of the Retractions
In August 2025 a team from Northwestern, Sydney, and the PNAS editorial board published something closer to a forensic audit than a research finding: a statistical trawl through who handled which manuscript at several journals, cross-referenced against which of those manuscripts were later retracted or flagged on PubPeer [1]. Nothing in the method involves reading anyone's mind or finding a confession. At PLOS ONE alone, of 18,329 editors who have ever accepted a submission, a Poisson-binomial enrichment test flags 45, a mere 0.25%, who together handled 1.3% of the journal's 276,956 articles but sit behind 30.2% of its 702 retractions. More than half of them, 25 of 45, also authored papers that PLOS ONE retracted. A tight cluster of them, spanning four countries, sent most of their submissions to each other rather than to the journal's other 18,284 editors. The same signature, self-referential acceptance, batched retractions, and a shared trail of duplicated images, reappears at Hindawi and across IEEE conference proceedings. This Field Note walks through exactly how the authors caught it, why a coordinated ring is a fundamentally harder catch than a single fabricated paper, and what the whole exercise implies, and does not imply, about trusting the other 99.75% of editors doing the job honestly. We do our own arithmetic on the paper's numbers throughout, clearly marked as ours; the club ran no experiment and analyzed no new data of its own.
MetascienceRead 24 minSections 14
Field notePaper Analysis
The Moth That Reads the Milky Way
Once a year, an Australian moth with a brain you could lose in a poppy seed flies up to 1,000 km to a handful of alpine caves it has never seen, packs in among roughly 16,000 other moths per square metre of cave wall for the summer, and then flies the whole distance back to breed and die. In 2025 a team led by David Dreyer and Eric Warrant tethered these moths inside a 50 cm planetarium, projected a real, moonless star field over their heads, and switched the Earth's magnetic field off with a bank of Helmholtz coils — nulled below 100 nanotesla, a reduction our own back-of-envelope arithmetic puts at somewhere north of 250-fold. The moths kept flying the correct seasonal bearing anyway. Rotate the projected sky 180° and the moths turn with it, in both directions, in both seasons (six trials, n from 49 to 70 each). Scramble the same stars into random positions and the moths go nowhere in particular. Recordings from the moths' own brains turned up 28 neurons, out of 79 sampled, tuned to the rotating sky, almost all of them firing hardest when the moth faced within a few degrees of due south — in spring and in autumn alike, headed toward the caves or away from them. This is, by the authors' own account, the first invertebrate ever shown to steer by the stars. We think that is a bigger deal than it sounds, and this note is our attempt to say exactly why, using nothing but the numbers in the paper and a calculator.
Animal NavigationRead 20 minSections 13
Field notePaper Analysis
The Six-Decade Blind Spot Above the Ozone Layer
Since Christian Junge found a persistent haze of sub-micron particles in the lower stratosphere with balloon-borne counters in 1961 [2], the layer has been treated as a sulfate story: droplets of sulfuric acid and water, replenished by volcanoes, sitting well inside the size range that satellites and balloon-borne optical particle counters can actually see. In April 2026, a team flying NASA's WB-57 high-altitude aircraft during the SABRE mission in February 2023 reported something that sulfate-only bookkeeping had never counted: a persistent mode of organic-rich particles, present below 19 km wherever the air still carries a nitrous oxide signature above 270 parts per billion, with a number-weighted mean diameter of only 0.03 to 0.11 µm [1]. Weighted by the surface area that actually does the chemistry, the same population runs 0.08 to 0.2 µm; weighted by volume, or mass, it runs 0.11 to 0.3 µm — three different "sizes" for one population, depending entirely on what you ask the size to explain. What it explains, per the paper's own accompanying institutional release, is up to 90% of the surface area available in the lower stratosphere for heterogeneous chemical reactions and for condensing trace gases, dwarfing the classic sulfate mode in the one property that actually drives ozone chemistry and particle growth. It went undetected for as long as it did because most satellite retrievals only resolve particles above roughly 0.2 µm, and balloon-borne optical counters generally start around 0.15 to 0.3 µm — both comfortably above where this mode's number distribution actually sits. A global chemistry-climate model run against the same data could not reproduce the population's properties, which the authors read as a warning for anyone using such models to project ozone recovery or to evaluate stratospheric geoengineering proposals that were designed, implicitly, around the sulfate picture alone. This Field Note works through why stratospheric surface area matters at all, does two pieces of arithmetic of our own directly on the paper's three reported size numbers, and asks plainly what changes and what does not.
Atmospheric ChemistryRead 20 minSections 14
Field noteExplainer
Two-Thirds of Cats Sleep on Their Left Side, Until Somebody Rechecked
In June 2025 a thirteen-author team led by Sevim Isparta coded 301 YouTube videos of sleeping cats and reported, in a two-page item in Current Biology, that 63.8% slept with their left side down — a population-level bias, they argued, that lines up with the right hemisphere's job of watching for threats through the left eye [1]. Eight months later, in a paper posted online in February 2026, one researcher working alone, Andrew Cooke of the University of Lincoln, ran the same question on three independent platforms and found nothing distinguishable from a coin flip anywhere: 51.2% left among 123 YouTube cats, 48.4% among 494 on Reddit, 48.8% average among 60 Instagram accounts [3]. Sitting between the two papers, easy to miss, is a formal correction: in October 2025 the original authors published an erratum removing 107 videos with data-quality problems, prompted, Cooke says, by concerns he had raised — and the corrected number came out exactly where it started [2]. This field note is about the pair, not either paper alone. It walks through what a “population-level bias” is actually claiming, why that is not the same statistical object as an individual “preference” the way pawedness is, how three sampling frames disagree with each other more informatively than any one of them disagrees with chance, and why a claim, a self-correction and a failed replication landing inside eleven months is what a functioning literature looks like doing its job. Every comparison across the two papers below — the pooled sample sizes, the percentage-point gaps, the chance expectations — is the club's own arithmetic on their published numbers, marked inline as ours.
MetascienceRead 18 minSections 13
Field noteExplainer
How Anybody Knows How Old Anything Is
Almost every date you have ever read about the past came out of one of seven or eight methods, and not one of those methods deserves to be trusted on its own. Tree rings and lake muds and ice layers count years directly, one at a time, while radiocarbon, potassium-argon and uranium-lead measure atoms instead, each covering a different slice of time. The whole structure becomes believable because the slices overlap, so any method can be checked against a second method that shares none of its physics, none of its instruments and none of its assumptions. We explain the decay statistics without calculus, simulate a population of atoms decaying one by one and watch the exponential law appear out of the randomness (0.051% agreement with the analytic curve over five half-lives, with no exponential anywhere in the code), and calibrate real radiocarbon ages against the real IntCal20 curve. Along the way: why 1% of modern contamination turns a genuinely ancient sample into a confident date of 36,993 BP, why a bigger sample stops helping past forty thousand years, and why a single measurement of 2450 ± 30 radiocarbon years maps onto three separate stretches of calendar time spread over 312 years. Also the argument against everything we just said, which is stronger than you might expect and which we have not tried to make easy on ourselves.
General ScienceRead 42 minSections 13
ResearchOriginal Research
Counting Every Word in Fifty Books to Test a Law Nobody Can Explain
Zipf's law says that if you rank the words of a text by frequency, the count at rank $r$ falls off like $r^{-a}$ with $a$ near 1. We tokenised fifty public-domain books from Project Gutenberg spanning fourteen languages and about 2,350 years, from a Greek tragedy to a 1925 American novel, and fit both a pure Zipf law and a shifted Zipf–Mandelbrot form by exact maximum likelihood, never by regression on a log-log plot, with goodness of fit assessed by the Clauset–Shalizi–Newman bootstrap. Across 4,714,004 tokens and 496,852 distinct word types, the rank exponent averaged 0.9549 (SD 0.0597) over all fifty books and 0.9899 (SD 0.0252) over the nineteen written in English, with every single book landing between 0.757 (a Finnish epic) and 1.045 (a Spanish novel). Splitting each book in half to measure real repeatability put the one-book measurement error at 0.0044, so the spread across books is 14 times larger than the noise: the differences are real, not sampling accidents. A pure power law was not rejected by the goodness-of-fit test in 34 of 50 books; a Zipf–Mandelbrot form with one extra free parameter beat it on AIC in all 50, and moved the fitted exponent by 13.8% on average, which is most of what a published Zipf exponent is worth without also stating which of the two forms produced it. Words that occur exactly once were undercounted by every fitted model, at a near-constant rate: the Zipf–Mandelbrot prediction averaged 45.8% of the true hapax count in every one of the fifty books. Validated against Clauset, Shalizi and Newman's own published fit of Moby-Dick, our exponent agreed to 0.29 combined standard errors. Validated against Moreno-Sánchez, Font-Clos and Corral's 31,075-text survey, our English-subset mean disagreed by −9.47 standard errors, a real and reported disagreement traced to tokeniser and corpus differences rather than to a coding error, demonstrated by rerunning our own pipeline under three other tokenisation rules and watching the exponent move by more than that gap. All five internal validations, including recovery of a known exponent from synthetic data and a from-scratch check of the Hurwitz zeta machinery against closed forms, passed before any real book was opened. Master seed 20260614. Corpus retrieved 2026-09-14 from gutenberg.org; all fifty downloads succeeded and none was substituted.
Summer 2026Read 33 minSections 12
ResearchMethods
When the First Digit Test Works and When It Accuses the Innocent
Benford's law says the leading digit of many collections of real numbers is 1 far more often than 9, following $P(d) = \log_{10}(1+1/d)$ , and it is now a standard test for cooked books, rigged elections and fabricated data. We put eight genuinely public datasets through the test, chosen for having nothing in common: country populations, CODATA physical constants, every FDIC-insured bank's total assets and total deposits, NASA's exoplanet orbital periods, a year of USGS earthquake magnitudes and the seismic moments those same earthquakes encode, and World Bank life expectancies. Two conformed closely (bank assets, MAD 0.0039). Two failed by a mean-absolute-deviation threshold while the chi-squared test called them unremarkable (populations, MAD 0.0152 against Nigrini's 0.015 cutoff, chi-squared $p = 0.47$ ). One passed the MAD band and still failed chi-squared at $p = 4\times10^{-11}$ (exoplanet periods). Two failed everything, badly, because they span less than half an order of magnitude (earthquake magnitudes and life expectancy, MAD 0.166 and 0.177). We then simulated log-uniform data across a controlled spread $S$ and found the deviation from Benford traces $|\sin(\pi S)/(\pi S)|$ almost exactly, landing on zero at every whole number of decades and peaking near half-integers, which is the mechanism the whole test rests on and the reason two of our eight datasets could never have passed regardless of anyone's honesty. Finally we ran the false-accusation arithmetic directly: 40,000 simulated honest Benford datasets per sample size showed the chi-squared test correctly calibrated at every $n$ (5.02% to 5.17% rejection against a nominal 5%, largest deviation 1.90 standard errors), while Nigrini's fixed MAD cutoff flagged 91.8% of honest datasets of size 100 and did not settle near its intended rate until $n$ reached the low thousands. Honest lognormal data with $\sigma < 0.3$ decades was rejected almost every time regardless of sample size. Five validations, including the closed-form digit law reproduced from first principles to 5.3 × 10⁻⁸, the club's chi-squared statistic against SciPy's to machine precision, and 2,000 powers of two checked against exact integer arithmetic, all passed. Seed 20260914. Every dataset was downloaded on 2026-09-14; none failed and none was substituted.
Summer 2026Read 31 minSections 12
ResearchMeta-analysis
Does Metabolism Really Scale to the Three Quarter Power
Kleiber's law says metabolic rate scales as body mass to the three quarter power. Rubner's older surface argument says two thirds. The gap between them is $\tfrac{1}{12} = 0.0833$ , and people have been arguing about it since 1932. We took one published compilation, AnAge Build 15, retrieved 14 September 2026, pulled out every species carrying both a body mass and a metabolic rate, and fitted the exponent by two methods. The sample is 627 species across four classes spanning 6.68 decades of mass, from a 0.76 g salamander to a 3,672 kg elephant. No animal was measured by us; the club has no laboratory, and the computation is the entire experiment. For mammals ( $n = 422$ ) ordinary least squares gives $b = 0.7091$ , 95% CI [0.6936, 0.7246], which excludes three quarters at 5.18 standard errors and two thirds at 5.37, and therefore rejects both. The same 422 species fitted by reduced major axis give 0.7273 [0.7120, 0.7430], still rejecting both, 2.6% steeper, because $b_{\mathrm{RMA}} = b_{\mathrm{OLS}}/|r|$ exactly. Pooled across all four classes the two methods land on opposite sides of three quarters: 0.7132 [0.6869, 0.7396] by least squares and 0.7883 [0.7623, 0.8151] by reduced major axis. Birds alone give 0.6549 [0.6358, 0.6739], consistent with two thirds and not with three quarters, and the mammal-bird slope difference of 0.0542 is 65% of the whole disputed gap ( $F_{1,589} = 10.38$ , $p = 0.0013$ ). Restricting the mass window is worse still: nineteen overlapping two-decade windows of mammals return exponents from 0.5911 to 0.8701, a range 3.35 times the quantity in dispute, and mammals under 100 g give 0.5959 while mammals over 316 g give 0.7959 with non-overlapping intervals. A quadratic term is strongly supported ( $t = 6.70$ , $p = 7 \times 10^{-11}$ ), so the local slope climbs from 0.572 at 10 g to 0.864 at 100 kg and no single exponent is the right answer to the question. Both regression routines were checked against exact rational closed-form solutions, agreeing to 1.8 × 10⁻¹⁵. Seed 20260614.
Summer 2026Read 32 minSections 12
Field noteMeta-analysis
How Much Carbon Can Farmland Actually Hold? A Review of the Field Trial Evidence
Farmers are being offered money to store carbon in their soil, the practices on the menu being no-till, cover cropping and residue retention, and the rates written into contracts usually sit somewhere near half a tonne of carbon per hectare per year. We went looking for where that number comes from. Screening 22 published syntheses and long-term trial reports, we could extract 12 effect sizes carrying usable uncertainty from six publications, then pooled them with a random-effects model we wrote ourselves and checked against synthetic data first. The pooled increase in soil organic carbon is +16.2% (95% CI 11.3 to 21.2), and nobody should use that number for anything, because $I^2 = 98\%$ and the 95% prediction interval for a new field runs from −1.7% to +37.3%. Two results survive the heterogeneity. Practices that import carbon onto the farm hold their effect below the plough layer while practices that only rearrange carbon already there lose it, and pooled, the two groups differ by a factor of 3.3 ( $p < 10^{-29}$ ). In a single dataset of 95 paired comparisons, switching from fixed-depth accounting to equal-soil-mass accounting cuts the measured no-till sequestration rate from 0.300 to 0.141 Mg C ha⁻¹ yr⁻¹, a factor of 2.1, with nothing whatever happening in the field [1]. Soils also saturate, and a one-line model fitted to no data and needing none says that a practice adding 0.5 Mg C ha⁻¹ yr⁻¹ of input buys about 53 years before the annual gain drops below a fifth of what it started at, and then stops. We are a school club reading abstracts rather than a systematic review with two independent screeners, and §3 says exactly what that costs.
Soil ScienceRead 38 minSections 13
Field notePaper Analysis
Ten Thousand Steps Was Never a Real Number
In 1965 a Japanese clock company named a pedometer manpo-kei — “10,000-step meter” — and sold it into the fitness wave building around that year's Tokyo Olympics. No trial produced that number. No physiologist chose it. It was a product name, and it outlived the product by sixty years, ending up as the default daily goal on phones and fitness trackers belonging to people who have never heard of the company that invented it. In August 2025, a team led by Melody Ding pooled 57 device-measured step-count studies — the broadest review of its kind to date, spanning mortality, cardiovascular disease, cancer, dementia, depression, diabetes, and falls in a single analysis — and found something more interesting than “the number is wrong.” Compared with 2,000 steps a day, walking 7,000 cuts all-cause mortality risk by 47%, dementia risk by 38%, and fall risk by 28% [1]. But the curves do not all behave the same way. Some bend hard around 5,000–7,000 steps and barely improve after that point. Others — cardiovascular mortality and cancer mortality among them — keep falling in something close to a straight line, with no plateau in the data at all. Seven thousand is not a magic number either. It is just the most honest one anyone has offered so far.
Public HealthRead 17 minSections 13
Field notePaper Analysis
What the Garum Vat Left Behind
Garum was Roman fish sauce, made by salting small fish and letting them rot on purpose, and it was made at industrial scale on the Atlantic coast of Iberia for centuries. What was left in the vats afterward was never meant to be looked at closely: crushed bone, digested by its own sauce, sitting for months in salt and acid and other creatures' enzymes. In 2025 a team working at Adro Vello, a Roman fish-salting site in Galicia, sequenced it anyway. From bone fragments in a vat dated to the first to third centuries AD, they recovered enough DNA to confirm the fish was European sardine, Sardina pilchardus, and to go further than any bone shape ever could: they placed those Roman-period sardines inside the species' modern population structure. Eighteen bone samples went in; seventeen came out sequenced, one discarded; endogenous sardine DNA ran from 4% to 21% of all reads, mitochondrial coverage from 1.3× to 197.6×. Set against 108 modern reference mitogenomes spanning the species' Atlantic and Mediterranean range, the Adro Vello sardines cluster with today's Central Atlantic group, Iberia and Morocco and the French coast, but noticeably less admixed than the modern population sampled from the same waters. This note is about how you pull a population history out of something a Roman worker threw away as sludge, and what a tighter, less-mixed sardine population implies about how a fishing fleet without engines actually worked its coast.
ArchaeogeneticsRead 20 minSections 14
Field notePaper Analysis
The Denisovans Finally Have a Face
For 92 years the most complete archaic human skull ever found in Asia had no scientific name, because a laborer in occupied Manchuria hid it down a well in 1933 rather than hand it to the wrong army, and his family kept the secret for three generations. By the time it reached a laboratory in 2018 and was described in 2021 as a new species, Homo longi, nobody could say what it actually was in genetic terms, because its bone held no usable DNA. In July 2025, Qiaomei Fu's team found the answer somewhere stranger: the hardened dental plaque on its teeth. Fu, Cao, Dai, Bennett, Feng, Yang, Ping, Pääbo & Ji reported mitochondrial DNA and roughly 95 endogenous proteins recovered from that calculus — the first time host DNA has ever been pulled from dental plaque on a Pleistocene human fossil — and both lines of evidence say the same thing: this is a Denisovan [1]. The mtDNA carries 27 diagnostic positions matching all seven previously sequenced Denisovan genomes and none of them modern human, which is exactly backwards from what museum contamination would produce. It is also the first time anyone has had a face to put on this population, previously known only from a finger bone, a few teeth, and two isolated jaws. We rebuild why dental calculus, of all things, turns out to protect biomolecules that bone loses, and we do our own arithmetic — not the paper's — on how deep the Denisovan mitochondrial branch really runs: about 1,000,000 years back to its common ancestor with ours, roughly 1.5× as old as the 660,000-year split between Neanderthal and modern human mtDNA, which is why 27 positions out of roughly 16,569 were all it took.
PaleogenomicsRead 20 minSections 13
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 $N$ -layer grey radiative-equilibrium system as a linear system for every $N$ from 0 to 10 reproduced the exact analytic answer $T_s = T_e (N+1)^{1/4}$ 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 $\varepsilon^\ast = \mathbf{0.7667 \pm 0.0122}$ 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
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
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
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 Q_between = 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 Q_M = 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
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
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
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
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
Field noteExplainer
What a P Value Actually Means, and the Four Things People Think It Means Instead
A p value measures your data against a pretend world, the one where the effect you are hunting does not exist. The number describes data, given a hypothesis. Swap those two around and every misreading in this article follows. We define the thing properly, then take apart four beliefs the club has had to argue with at its own table: that p is the probability the hypothesis is wrong, that a smaller p means a bigger effect, that 0.049 and 0.051 are different animals, and that a non-significant result means nothing happened. Our own simulation of 200,000 experiments under a true null backs the definition, with the p values coming out flat to a Kolmogorov-Smirnov distance of 0.0014 from perfect uniformity. Another 400,000 simulate a whole literature. Now the headline. In a field where 10% of tested hypotheses are true and studies have 50% power, 47.4% of everything published as significant at $p<0.05$ is false, with no fraud, no p-hacking, no hidden analyses and no arithmetic errors anywhere. Two honest studies of one real effect disagree half the time, and one line of arithmetic gives you both numbers. No results section reads the same way afterwards.
StatisticsRead 29 minSections 13
ResearchMethods
The Treatment That Works Because You Picked the Worst Cases
Take a population, measure everybody, keep the worst tenth, do absolutely nothing to them, and measure them again. In our simulation they improved by 0.8760 ± 0.0013 standard deviations. The instrument was a program the club wrote, so the computation is the experiment and no person, patient or pupil entered the study at any point. Across a grid of six selection thresholds ( $p = 0.50$ down to $0.01$ ) and seven noise-to-signal ratios we simulated 372 million individuals, each measured twice, with a stable underlying trait and fresh noise on each occasion. Every cell was checked against the closed form. For a bivariate normal with correlation $r$ , a group selected on the first measurement has its mean shrunk toward the population mean by exactly $r$ , and our measured shrinkage matched $r$ in all 42 cells, the largest discrepancy being 2.53 standard errors against a Bonferroni threshold of 3.41; the pooled fit gave $\sum z^2 = 36.20$ on 36 degrees of freedom, $p = 0.459$ . With zero noise, where $r = 1$ , the apparent improvement was identically zero at every threshold, which is the control that makes the rest of the grid mean anything. Adding a genuine treatment effect, a single-arm before-and-after design overstated a true 0.20 SD effect by 5.38× in our headline cell, attributing 81.5% of its result to an intervention that supplied none of it, while a randomised design run on the same simulated people recovered the truth to within 0.87 standard errors. At a true effect of zero the single-arm design reached $p < 0.05$ in 100% of replicates. Set against the published figure of $-0.24$ for change from baseline in the untreated arms of 37 three-armed trials [15], regression alone accounts for the whole of it once entry reliability falls to 0.863 at a bottom-tenth cut. Seed 20260211.
Winter 2026Read 31 minSections 12
ResearchMethods
Building a Dataset Where Every Group Trends Up and the Whole Trends Down
Simpson's paradox is usually shown once, with one famous table, and then left alone as a curiosity. We treated it as a piece of engineering instead. The club wrote a generator that takes a target within-group slope and a target pooled slope of the opposite sign and returns a dataset hitting both, by solving for the required between-group slope rather than searching for one. Across eight specifications the worst deviation between the fitted slopes and the requested ones was 4.44×10⁻¹⁶. Nothing was measured in the world; the computation is the experiment. Then we ran the machine backwards and asked how often the paradox arrives uninvited. With four groups of sixty points, a within-group slope of $+1$ , and a group-level confounder as strong as the within-group spread in $x$ , 9.70% ± 0.21% of 20,000 simulated datasets showed every group sloping one way and the pooled fit sloping the other, and 6.04% ± 0.17% did so with every slope involved clearing a two-sided $t$ test at $\alpha = 0.05$ . Turning the confounder off gave 0 reversals in 40,000 datasets, which it must, because the pooled slope is then equal to the within-group slope as an algebraic identity. Group-size imbalance turned out to work the opposite way from the folklore, cutting the rate from 9.70% to 4.86% at a 30:1 ratio, and coupling size to the confounder cut it further to 2.43%, a difference of 13.0 standard errors. In the continuous case with a single lurking variable the Monte Carlo matched a closed-form integral to within 1.60 standard errors at every strength. As an external check, our 3,000,000 random 2×2×2 tables reversed at 0.016734 ± 0.000074 against Pavlides and Perlman's exact 1/60 [10], a gap of 0.90 standard errors. Seed 20260214.
Winter 2026Read 30 minSections 12
ResearchOriginal Research
How Many Reasonable Choices Does It Take to Find a Result That Is Not There
Two groups of numbers drawn from the same distribution cannot differ. We drew them 50,000 times and then analysed them the way data actually gets analysed, making one defensible choice after another with the numbers already in front of us. With no choices at all the false positive rate came out at 0.0504 ± 0.0010, a 95% interval of 0.0486 to 0.0524 that covers the nominal 0.05, and that single number is the check which licenses everything else here. Five researcher degrees of freedom were then applied singly and in all 32 combinations: peeking four extra times while enrolment ran from 20 to 40 per group, testing three correlated outcome measures instead of one, splitting on an irrelevant binary attribute, adding or omitting an uninformative covariate, and excluding observations beyond two standard deviations. Singly the most expensive was outcome choice, at 0.1286 ± 0.0015. Together the five produced a significant result in 35,331 of 50,000 studies of pure noise, a rate of 0.7066 ± 0.0020, which is 14.1× the nominal level. The 180 analysis paths this permits behave like 23.9 ± 0.1 independent tests rather than 180, because they share data; 180 genuinely independent tests would give 0.99990. The club's t statistic agrees with scipy.stats.ttest_ind to 1.3×10⁻¹⁵ on identical data, and the covariate-adjusted statistic agrees with a LAPACK least-squares solve to 3.6×10⁻¹⁵. Optional stopping alone, with equally spaced looks, reproduces the 1969 inflation table of Armitage, McPherson and Rowe at every published entry, the largest deviation being 2.7 standard errors at twenty looks. Three situations from Table 1 of Simmons, Nelson and Simonsohn (2011) were rebuilt to that table's own specification and landed within 2.7 standard errors of the published values at all three significance levels. The headline depends on choices we made and could have made otherwise: at outcome correlation 0.8 rather than 0.5 it falls to 0.613, and under a three standard deviation outlier rule rather than two it falls to 0.573. Seed 20260214.
Winter 2026Read 34 minSections 12
Field noteData Story
An Empire's Air Pollution, Measured in Greenland Ice and in Roman Children's Blood
Three Arctic ice cores hold a dated record of the smoke that European silver smelters put into the air two thousand years ago. In January 2025 a team led by Joseph McConnell at the Desert Research Institute ran that record backwards through an atmospheric transport model to recover annual emissions, then forwards again to map lead concentrations across Europe, then through two published epidemiological regressions to reach the blood of Roman children [1]. Their emission figure for the Pax Romana is roughly 3 to 4 kilotons of lead a year. Their arithmetic carries that to a childhood blood lead level about 2.4 micrograms per decilitre above background, and from there to an empire-wide cognitive cost of 2.5 to 3 IQ points. We rebuilt the chain. We replaced their dispersion model with a single well-mixed box over Europe and left the two published regressions untouched. Out came 0.514 nanograms of lead per cubic metre of air, a blood lead rise of 2.39 µg/dL, and a decrement of 2.61 IQ points. Close enough to examine. Not close enough to trust blindly. Run the whole chain across the plausible parameter space and something else appears. The atmospheric physics, the hard part, barely matters. A twelvefold change in emissions moves the final answer by 27 per cent. A one-standard-error wobble in a single epidemiological exponent moves it by 57 per cent.
Environmental ArchaeologyRead 29 minSections 12
Field notePaper Analysis
A Volcano That Has Been Dead for 250,000 Years Is Still Breathing
Uturuncu is a 6,008-metre dacite volcano in southwest Bolivia whose last eruption was 250,000 years ago, give or take five thousand. For three decades it has impersonated a mountain about to go off. The ground above it has been rising at roughly a centimetre a year over a broad dome, that dome sits inside a ring of slow subsidence reaching past 70 kilometres, and small earthquakes fire underneath it most days. In April 2025 a team led by Ying Liu and Mike Kendall published a joint seismological and petrophysical anatomy of its interior, built from more than 1,700 local earthquakes, and concluded that the unrest is gas and brine climbing a narrow chimney out of a much older magma body rather than fresh magma on its way up [1]. We rebuilt the deformation half of the argument ourselves. Working on a synthetic sombrero pattern of our own making, we show that no single pressurised point source at any depth can produce a ring of subsidence around a central uplift, because the vertical displacement of such a source carries the sign of its volume change at every radius without exception. The best single source leaves a misfit of 1.29 mm/yr, about six and a half times the noise we put in; a pair at 20.8 and 77.4 km gets to 0.17 mm/yr. Then we divide. The fitted inflation rate of 2.5 × 10⁷ m³/yr is 192 times Uturuncu's own long-run magma output rate, and had it run steadily since the last eruption it would have delivered 6,360 km³, seventy-five times the volume of the entire mountain. Both calculations are the club's own and both are crude. The direction of the conclusion does not depend on either.
VolcanologyRead 39 minSections 12
Field notePaper Analysis
Twenty-Three Minutes of Chewing, One Kilometre Up
Carlos Ibáñez found the first clue in 2001: feathers, too many to be nesting material, packed into the guano beneath a greater noctule bat roost in southern Spain [2]. In 2007 his group ran the blood chemistry and got an isotope signature that, every autumn, sat inside the zone reserved for animals eating almost nothing but birds [3] — a gap of 3.21‰ in carbon and 3.15‰ in nitrogen from the bats’ ordinary insect diet, our own subtraction of the paper’s reported means. Neither result could say how a bird ends up eaten by something that echolocates. Caught alive, a kilometre up, by an animal the size of a fist? Or found already dead? Isotopes cannot tell a kill from a carcass; the molecules read the same either way. For eighteen more years nobody could resolve it, because nobody could follow a bat far enough into the dark to watch. In October 2025 a team led by Laura Stidsholt did, by making the bat carry the witness. Sound-and-movement tags built at Aarhus University, riding on Nyctalus lasiopterus over Spain’s Doñana reserve, recorded two hunts start to finish: a 30-second chase that failed, and a chase lasting nearly three minutes — more than a kilometre above the ground, in total darkness — that ended with 21 recorded distress calls from the bird, a robin, followed by 23 minutes of chewing sounds as the bat kept flying and ate it in the air [1]. This field note follows the case from the guano pile to the recording, and asks the question the tag actually answers: what does a hunting technique built for insects a few centimetres long look like when it is redirected at something built to survive a hawk?
Animal BehaviourRead 16 minSections 13
Field notePaper Analysis
Killer Whales Are Manufacturing Grooming Tools, Mostly With Family
In the spring and summer of 2024, a drone hovering over the central Salish Sea filmed something nobody had described before: a southern resident killer whale biting a short length of stipe off a whole bull kelp stalk, pressing the piece between its own body and a partner's, and rolling it back and forth for as long as 12 minutes and 4 seconds. Michael Weiss, Darren Croft and six coauthors call the behavior allokelping, and their June 2025 report in Current Biology documents 30 bouts of it across roughly 10 hours 21 minutes of footage, spanning all three southern resident pods [1]. The kelp itself is nothing special. What earns the paper its attention is the verb: on seven of eight clearly filmed acquisitions, the whale did not pick up a loose scrap, it bit one to length off a still-rooted stalk, the "detach" mode in the standard taxonomy of animal tool manufacture [2]. This field note works through why that distinction is the whole story, what a Bayesian model of who-grooms-whom in one pod actually found — maternal kinship predicts a pairing with posterior probability >0.99, similar age with probability 0.98 — and does the club's own long division on the paper's own reported effort and bout counts. We add no data of our own. We only divide what the authors already published, and we say so every time we do it.
Animal BehaviorRead 20 minSections 13
ResearchOriginal Research
Why Sex Exists: Watching Mutations Pile Up in a Population That Cannot Recombine
Mutation only adds. In a population with no recombination, once the last individual carrying the fewest mutations fails to leave descendants, nothing can rebuild that class, and the minimum mutation load in the population steps up by one forever. Muller called it a ratchet. We measured how fast it turns. The instrument was a Wright-Fisher simulation the club wrote and ran on a laptop, carrying the full distribution of mutation counts per individual; no organism was involved at any stage, because here the computation is the experiment. Across a 60-cell grid ( $N = 100$ to $2000$ , $U = 0.1$ to $0.8$ , $s = 0.02$ to $0.1$ ) plus 35 further cells, at 200 replicate populations each, we logged 11,220,000 replicate-generations in the main grid alone. The null passed exactly: with $U = 0$ the ratchet clicked zero times in 1,200,000 replicate-generations and mean fitness stayed at 1.000000000000000. Haigh's deterministic equilibrium survives where it is defined, the mean load matching $\lambda = U/s$ to +2.9% and the size of the unloaded class matching $N e^{-\lambda}$ to −4.7% ± 1.1%. The rate prediction built on it does not survive. Across 70 clicking cells the measured rate was a median 0.398× the classical $e^{-n_0}$ , and in the worst cell it was 1.4×10⁵⁵ times larger, sitting 11 standard errors above a prediction of $1.7\times10^{-59}$ . Three independently written implementations agree on our number to within 1.9 standard errors, so the fault is in the formula. The missing quantity is the selection coefficient: at fixed $n_0 = 5$ raising $s$ from 0.02 to 0.10 nearly doubled the click rate, a gap of 28.5 standard errors, and rescaling by $s \times n_0$ collapsed cells that differ five-fold in both onto one curve within 6.9%, against 78.4% disagreement when matched on $n_0$ alone. Mean fitness fell exactly as bookkeeping demands, the measured $d\ln \bar{W}/dt$ matching (click rate) $\times \ln(1-s)$ to 0.9961 ± 0.0031 over six regimes. Adding recombination to an explicit 1024-site genome stalled the ratchet: 2.7% of offspring made sexually halved it, and 15.2%, roughly one offspring in seven, cut it to a hundredth of the asexual control. Seed 20250921.
Fall 2025Read 34 minSections 13
ResearchOriginal Research
The Dosing Schedule Decides Whether Resistance Wins
We built a stochastic two-strain bacterial population model under a one-compartment pharmacokinetic drug concentration and ran 178,002 simulated treatment courses on a laptop. There are no bacteria in this study and no patients: the computation is the experiment, and none of what follows is medical advice. Before measuring anything we validated the deterministic core. Bisection on the net growth rates recovered the mutant selection window as [1.000000000000, 16.000000000000] mg/L against the input MICs, agreeing to 2.9 × 10⁻¹⁵ mg/L; the lower edge of the wider selective window matched a closed form we solved by hand to 2.7 × 10⁻¹⁵ mg/L, landing at MIC/5.26, inside the MIC/4 to MIC/230 range Gullberg and colleagues measured experimentally [8]. With the fitness cost set to zero the resistant strain never declined without drug (minimum net growth rate exactly 0.000000/h); with mutation switched off, resistance appeared in 0 of 2000 replicates, and a schedule that cures only 8.90% of the time once mutation is switched back on cured every one of them. Across a grid of five dose sizes and five dosing intervals at perfect adherence, peak concentration relative to MIC was the best single predictor of cure (Spearman ρ = +0.9809) and total drug delivered was worse (+0.9333). At matched total exposure of 1246.5 mg/L·h, a two-day schedule of 48 mg/L every 8 h cleared 2000 of 2000 replicates while a fourteen-day schedule of 6.857 mg/L delivering exactly the same drug ended in resistance 91.10% ± 0.64% of the time, a gap of 143 standard errors. Holding dose fixed instead, one day of treatment cured 86.70% ± 0.76% against 100% for fourteen days, and the median time to clearance was 6.0 h in every row. Missing doses mattered, taking cure from 1.0000 to 0.7670 ± 0.0094 at 50% adherence, but a control that delivered the same reduced total drug with no gaps did considerably worse (0.4980 against 0.8595 at 80%). The short schedule won in all fourteen sensitivity variants, with the margin collapsing from +0.880 to +0.022 at a hundred-fold smaller inoculum. Seed 20250915.
Fall 2025Read 35 minSections 12
ResearchReplication
Running Axelrod's Tournament Again, With Noise This Time
Axelrod's 1980 round robin found that tit for tat beat every program sent to it. We asked whether that survives when moves are occasionally misread. No person and no organism was involved at any point: the computation is the experiment. We implemented fifteen classical strategies, ran a full round robin of 400 repetitions at each of ten noise levels from 0 to 0.20, and then ran replicator dynamics for 4,000 generations on each measured payoff matrix. The main sweep is 480,000 matches and 96 million rounds. Fifty analytic pairwise payoffs at zero noise were reproduced exactly, the payoff structure satisfies $T > R > P > S$ and $2R > T+S$ , and the replicator update conserved total frequency to $5.55 \times 10^{-16}$ over every generation of every run. At zero noise tit for tat scored 2.62890 and finished third: tied with contrite tit for tat (paired $t = 0.47$ ) and 3.04 standard errors behind generous tit for tat at 2.63999. A leave-one-out sweep shows the entire margin comes from one opponent, Joss. Under noise the board turns over twice. Generous tit for tat leads to 2% noise, contrite tit for tat from 5% to 15%, and at 20% noise always defect wins with 2.25965. At 5% noise tit for tat trails the leader by 0.07785 per round, a paired $t$ of 24.1. The evolutionary run is cleaner: contrite tit for tat takes the entire population at every noise level from 0.005 upward, while the mean payoff of that all-cooperative population falls from 3.0000 to 2.6421, a loss of 11.9% caused by errors alone. An exact Markov chain for two tit-for-tats agrees with the simulation to within 1.84 standard errors at every noise level. Seed 20250913.
Fall 2025Read 31 minSections 12
Field noteExplainer
Snowball Earth Was Flickering: A Planet Stuck in a 56 Million Year Loop
The Sturtian glaciation opened around 717 million years ago and closed around 661 million years ago, and the 56 million years of ice lying between those two dates have embarrassed every account of the Cryogenian written since anyone first learned to measure the interval at all [5, 6]. A Harvard group published a different reading of it on 27 April 2026 [1]. In their coupled model of Neoproterozoic climate and carbon, the rapid weathering of the newly erupted Franklin basalt province destroys every stable climate the planet has available to it, so that Earth never rests anywhere but circles instead, frozen and thawed and frozen again, turn after turn, across the whole 56 million years. We rebuilt the mechanism in a deliberately small model of our own, two equations rather than a research code, because we wanted to watch the thing work. Our version puts the freeze threshold at 257 ppmv of carbon dioxide and the thaw threshold at 30,328 ppmv, a factor of 118 apart, and finds that any weatherability above 3.59 times the modern value leaves the planet nowhere stable to sit. Out of that comes a cycle with a period of 1.31 million years and a peak-to-trough swing of 68.8 K, spending 94% of each turn frozen: roughly 43 freeze-and-thaw cycles inside the Sturtian. The digits belong to us rather than to the paper, and the mechanism matters far more than the digits do.
PaleoclimateRead 33 minSections 12
Field noteData Story
Twenty One Percent of the Ocean Got Darker and Almost Nobody Noticed
Two researchers at Plymouth took twenty years of satellite measurements of how fast light dies in seawater, one number per pixel per year on a 9 km grid, and asked a question nobody had asked globally: is the ocean getting dimmer? Over 2003 to 2022 the attenuation coefficient rose across 75,341,181 km², which is 21 percent of the ocean, while it fell across only 10 percent. The lit layer shoaled by more than 50 metres over 32,449,129 km² and by more than 10 metres over 19 percent of the sea [1]. We rebuild the underlying optics from the Beer-Lambert equation, work out why clear water loses hundreds of times more depth per unit of added murk than muddy water does (a ratio of about 2,010 to 1 between the clearest gyre and a river plume), build a Secchi disk lookup table anyone with a rope can use, and then add up the lit water that disappeared: 3.75 million km³, or about one Mediterranean Sea, roughly 0.28 percent of all the water in the ocean. Our volume figure is a bookkeeping exercise built on the paper's reported areas, not a measurement, and the section on objections explains why twenty years may still be too few.
OceanographyRead 27 minSections 12
Field notePaper Analysis
The Slug That Eats Its Own Solar Panels
Some sea slugs eat algae, digest everything except the chloroplasts, and keep those chloroplasts running inside their own cells — photosynthesizing, on borrowed hardware, for nearly a year, a state biologists call kleptoplasty. Nobody knew how an animal cell kept a plant organelle alive that long without the plant's own nucleus to keep supplying it with parts. In June 2025 a team led by Corey Allard and Nicholas Bellono answered it: the slug Elysia crispata builds a dedicated compartment, wraps one stolen chloroplast in it at a time, and marks the membrane with its own late-endosome machinery — proteins named Rab7, VHA, P2X4 and NPC2, all host-encoded, none algal [1]. They named the compartment the kleptosome. An ATP-gated channel in its membrane, P2X4, tunes the interior to keep photosynthesis running; block that one channel with the antagonist 5-BDBD and net oxygen output falls 62% (p = 0.0250) [1]. Then comes the part that reframes the whole system: starve the slug, and it digests the chloroplasts stored inside its own kleptosomes for calories. Elysia crispata survives close to four months without food this way, next to three to four weeks for a non-photosynthetic relative, the sea hare Aplysia californica, run through the identical experiment [1]. A structure built to keep a solar panel switched on turns out to double as the battery it drains when the lights would otherwise go out. The same authors found a convergent version of the trick in coral and anemone tissue, which arrest and digest their own symbiotic algae by a similar route [1]. This field note is our own read of a paper we did not write and an organism we have never touched; we did not run a lab, and every number below traces back to Allard et al. or to the sacoglossan literature they cite.
Cell BiologyRead 18 minSections 13
Field notePaper Analysis
The Trap That Evolution Tried Exactly Once
In March 2025 a team led by Qiong Wu and Lars Vilhelmsen described a wasp trapped in 99-million-year-old Burmese amber that does something no living wasp does, and, as far as anyone can tell, no fossil wasp ever did again [1]. Sixteen females of †Sirenobethylus charybdis carry, at the very tip of the abdomen, a hinged three-flap apparatus — one plate each from a modified sternum 6, tergum 7 and sternum 7 — fringed with long trigger setae and caught, across the sixteen specimens, in a spread of open and closed positions. The authors call it, cautiously, reminiscent of a Venus flytrap, and write that nothing similar is known from any other insect, living or fossil [1]. This note is not about the wasp finding a host. It is about how you prove an abdomen used to grab things when the abdomen has been dead for 99 million years and cannot be watched doing anything. We walk through the comparative anatomy the authors leaned on, the muscle-attachment reasoning built from two unusually large sclerites with a straight line across them, and the process of elimination that ruled out the one alternative hypothesis anyone could think of. We also do a little arithmetic of our own, none of it the paper's: the amber is dated to 98.79 ± 0.62 million years [17], which places this wasp roughly 33 million years before the asteroid that ended the Cretaceous [20] — meaning the entire 66-million-year run of the Cenozoic happened after this structure and nothing living inherited it. A trait does not need to fail to disappear. That is the part of the story worth sitting with.
PaleoentomologyRead 23 minSections 13
Field notePaper Analysis
The Phase Transition Hiding in Your Cacio e Pepe
Cacio e pepe is three ingredients — pecorino, black pepper, pasta water — and it fails constantly, seizing into a grainy, stringy mess that any home cook has produced at least once. In April 2025, eight physicists gave that failure a name and a boundary. Writing in Physics of Fluids, Bartolucci and colleagues showed that the clumped state is a real thermodynamic phase, which they call the Mozzarella Phase, separated from a smooth one by a starch concentration of almost exactly 1% relative to the cheese's mass [1]. Below that line the sauce clumps once it passes roughly 65 °C; above it, starch holds the cheese proteins apart all the way to 95 °C. The transition is a lower critical solution temperature, sitting slightly above 60 °C at a protein mass fraction of φ ≈ 0.134, and the physics runs backward from kitchen folklore: it is heating the mixture too far, not letting it cool, that breaks it. This note works through how eight scientists turned a ruined dinner into 50–95 °C temperature ramps over backlit Petri dishes, and does the club's own arithmetic on the recipe they end up recommending: 5 g of starch against 200 g of cheese is 2.5% by mass, about 1⅞ teaspoons by a standard kitchen conversion, and it puts the actual dish at a protein fraction of φ ≈ 0.154 — close enough to the paper's most clump-prone concentration that the 1% line is not a formality.
Soft Matter PhysicsRead 22 minSections 13
ResearchOriginal Research
Spots, Stripes and the Two Chemicals That Might Explain Both
In 1952 Alan Turing proposed that two substances reacting and diffusing at different rates could turn a uniform sheet into a patterned one with no template anywhere. We tested the mechanism by integrating the Gray-Scott system on a periodic 256 × 256 grid, and nothing here was observed in an organism or a beaker, because the club has no laboratory and the computation is the experiment. We derived the Turing conditions in closed form and evaluated them at $F = 0.0740$ , $k = 0.0620$ , $D_u/D_v = 2$ , where the uniform state is stable without diffusion ( $\mathrm{tr}\,J = -0.014209$ , $\det J = +0.000300$ ) and unstable with it ( $B = +0.009743 > 0$ ). Growth rates measured mode by mode matched the analysis to a worst relative error of 1.08 × 10⁻⁵ over twelve modes, and during growth the pattern's wavelength came out at 13.6371 ± 0.1372 cells against a predicted 13.5612, a gap of +0.56% and z = +0.55 over 24 replicates. After the pattern stopped moving it had drifted to 13.8870 ± 0.0877 cells against the fastest-mode prediction of 12.7207, +9.17% and z = +13.30; a 512 × 512 control put the same number at 13.8186 ± 0.1150, ruling out the grid. Setting $D_u = D_v$ annihilated the pattern exactly as the theory requires: the surviving amplitude after 3000 steps was 6.47 × 10⁻¹⁴ against 0.1553 for the unequal case, and the measured decay rates sat within 2.78% of the analytic ones. A 169-point sweep of $(F, k)$ produced 123 uniform, 14 spots, 18 maze, 14 chaotic and zero aligned-stripe outcomes; stripes are stable when seeded but are never chosen out of isotropic noise. Wavelength fell with diffusion ratio as $r^{-0.678 \pm 0.036}$ against a predicted $r^{-0.687}$ , a gap of 0.26 standard errors, and at the low end of that sweep the pattern grew exactly on schedule and then collapsed to nothing at all. Seed 20250621.
Summer 2025Read 35 minSections 13
ResearchOriginal Research
Why Frost, Coral and Copper Deposits All Grow the Same Shape
Branched, feathery growth turns up in places that share almost no physics, and the club wanted to know whether the family resemblance survives being measured rather than merely admired. We implemented diffusion-limited aggregation on a square lattice and grew six independent clusters of 50,000 particles, plus 192,000 further particles in a sticking-probability sweep and an algorithmic control, for 492,000 particles in total. Nothing here was observed, because the computation is itself the experiment, and two estimators turned loose on the same six clusters were asked to hand back the same number. Radius-of-gyration scaling gave D = 1.6992 ± 0.0172, which sits 0.63 of our own standard errors below the accepted value of 1.71, while box counting on those identical clusters gave D = 1.6420 ± 0.0015, which is 46.9 standard errors below it. The two rulers therefore disagree with each other by +0.0572 ± 0.0172, a paired z of +3.32, and we spent most of this study on that gap rather than on the number that flattered us. Running that same box counter on three shapes whose dimension is known exactly showed it reading low in every case, by −0.119 on average. The radius-of-gyration exponent refuses to hold still either: its local value falls from 1.824 at a few hundred particles to 1.669 over the last half-decade, while a four-fold anisotropy grows from 0.041 to 0.416, the known signature of the square lattice rather than of the growth process. Dropping the sticking probability from 1.00 to 0.05 raised the apparent dimension at fixed size from 1.7006 ± 0.0259 to 1.9595 ± 0.0171 and shrank the mean radius of gyration by 43.6%. An algorithmic control found no detectable bias from our long-jump acceleration, −0.0251 ± 0.0262. Seed 20250704, runtime 365 s.
Summer 2025Read 36 minSections 12
ResearchReplication
A Planet of Daisies That Regulates Its Own Temperature
In 1983 Andrew Watson and James Lovelock published a planet made of black and white daisies, to answer the charge that Gaia required the biosphere to plan ahead [3]. We re-implemented their model, swept the solar luminosity from 0.6 to 1.6 times present in 501 steps, integrated to steady state at each, and observed nothing whatever, since the computation is the experiment and no planet, daisy, or physical measurement appears anywhere in this study. The daisies hold the planetary temperature inside a band of 10.421 °C across L = 0.704 to 1.556, which is 85.2 per cent of the swept range, while an identical bare planet moves through 60.236 °C over the same interval: a suppression factor of 5.78, rising to 8.30 across the middle 90 per cent of the interval and 6.65 inside the band where temperature stays within 2 °C of the growth optimum. Three checks make this a measurement rather than a picture. The three surface areas, integrated as independent variables, sum to one to within 8.882×10⁻¹⁵, or 40 machine epsilons, at every one of 1,162,915 integration steps, and the daisy-free control reproduces the Stefan–Boltzmann temperature exactly, to 0.000×10⁰ K at all 501 luminosities. Setting both daisy albedos equal to bare ground, which leaves the ecology untouched and removes only the radiative channel, collapses the temperature curve onto the bare control to 2.842×10⁻¹³ K while daisies still cover up to 70 per cent of the surface. An independent closed-form prediction we derived and did not impose, that the bare-ground fraction is pinned at γ/(1−(d/h)²) = 0.326530786 throughout the coexistence band, is reproduced by the integrator to 1.7×10⁻⁹. A Monte Carlo over 2,000 randomly drawn parameter sets gives a regulated width of 0.8066 ± 0.0025 and a suppression factor of 6.403 ± 0.051, with every drawn planet regulating to some degree. Against that, one bookkeeping choice we did not measure, whether dead daisies can re-seed, moves the hot edge of the regulated range from 1.556 to 1.212 and removes 40.4 per cent of the result. Seed 20250621.
Summer 2025Read 36 minSections 12
Field notePaper Analysis
The Tree That Wants to Be Struck by Lightning
For a century, lightning has been filed under damage. In March 2025 a team working in central Panama filed it under strategy. A custom lightning-location array pinned down 93 struck trees, and the crews then went and looked at every one of them, which is how ten strikes on Dipteryx oleifera, the tonka bean, came to be recorded as doing negligible damage while 64% of struck trees of other species were dead within two years. The same bolts killed 78% of the vines strangling the Dipteryx crowns, along with roughly 2.1 tonnes of neighbouring competitor biomass per strike, and the authors estimate that surviving lightning multiplies lifetime seed output about 14-fold [1]. This Field Note asks something the paper does not. Does the physics even permit that asymmetry? We build our own crude two-resistor model of a struck trunk and its lianas, the club's own simplified model rather than the paper's analysis, and the answer turns out to depend violently on one modelling choice. Share a voltage between trunk and vine and the vine heats 0.75× as fast per kilogram as the trunk, so nothing burns anywhere; let the vine carry the current itself and it heats 271,205× faster, boiling dry in 243 microseconds. Our 300-year Monte Carlo returns a fitness advantage of 7.23×. Large and real, and still only half the paper's 14×. We show exactly which assumptions we would have to bend to close that gap. We do not bend them.
Tropical EcologyRead 33 minSections 12
Field notePaper Analysis
Clownfish Shrink to Survive Heatwaves, Actually Shorter, Not Just Thinner
A bony vertebrate is supposed to get longer, or hold still. Reversing is not on the menu. During the marine heatwave that hit Kimbe Bay, Papua New Guinea in 2023, Melissa Versteeg and colleagues measured 134 wild clown anemonefish six times over five lunar months. A hundred of them got shorter. Shrinking raised an individual's survival probability by as much as 78 percent. Pairs where both partners shrank did better than pairs where only one did [1]. We rebuilt the physiological argument ourselves, from oxygen solubility tables and scaling exponents, and the club did not expect what came out. Under the gill-limitation account, a typical 1.4 percent shrink offsets 0.0248 °C of warming, against a 4 °C anomaly. Change the accounting to count total energy demand instead of supply per unit demand, and the same shrink is worth 0.4577 °C, eighteen times more. Our simulated survival model recovers an odds ratio of 1.73, and tandem shrinking beats solo shrinking by 7.65 percentage points. Everything past §5 is the club's own simplified calculation, kept clearly separate from what Versteeg and colleagues measured, and one of its inputs can flip the sign of the whole geometric argument.
Marine PhysiologyRead 33 minSections 12
Field notePaper Analysis
Eight Thousand Kilometres on a Raft of Weeds: How Iguanas Reached Fiji
Fiji has iguanas. Nothing else in the Pacific does, and the family is otherwise an American one, so for a century the four Fijian species were explained away as leftovers from a vanished southern supercontinent or as the end of a long chain of island hops from South America. In March 2025 a team led by Simon Scarpetta sequenced more than four thousand genetic loci and found the closest living relative of the Fijian iguanas sitting in the Mojave and the Sonoran, the desert iguana, with the two lineages parting somewhere near 34 million years ago [1]. The implied journey is a drift of more than 8,000 km across open ocean, the longest crossing ever documented for a land vertebrate, and we wanted to know whether an animal could physically survive it, so we built our own water and energy budget for a 70 g lizard on a floating mat of vegetation. Our simplified model gives 179 to 833 days at sea depending on the current, a fasting limit of 73.7 days, and a dehydration limit of 27.3 days with no fresh water at all, both of them far shorter than the voyage. The trip only closes if the raft is edible and it rains. On our numbers it needs rain on roughly 15% of days to give a coin-flip chance of arrival, which the tropical Pacific supplies easily along the whole route, then and now. Everything after the second section is the club's own arithmetic rather than the paper's analysis, and one of our parameters could move the answer by two orders of magnitude, which we name and flag rather than bury.
BiogeographyRead 34 minSections 12
Field notePaper Analysis
What Nine Egyptian Mummies Actually Smell Like
A smell has never had a fixed procedure. In February 2025 a team led by Emma Paolin, working across University College London, the University of Ljubljana, Kraków University of Economics, and the Egyptian Museum in Cairo, gave it one anyway. They put nine mummified human bodies — five held in storage, four standing in gallery cases, spanning the New Kingdom to the Byzantine period — through non-invasive headspace sampling, gas chromatography-mass spectrometry with olfactometry, and a trained sensory panel scoring thirteen fixed descriptor words [1]. The three words that came back most often were woody (in 78 percent of the bodies), spicy (67 percent), and sweet (56 percent) — a chemical signature of cedar and juniper resin, myrrh, and beeswax that has been sitting in these bodies for roughly three thousand years. Underneath that signature the instrument also found compounds nobody embalmed on purpose: a botanical pest-control oil the museum itself applies to display cases, unexplained synthetic pesticide residues, and volatile by-products of active mould and bacteria still living on the wrappings. This field note is a close read of the method — how you turn a subjective sensation into a chemical identity two people can agree on — the compound list, and the paper's own ethics section on what it means to study, and publish, the smell of a human body that cannot consent to being smelled. The arithmetic converting one of the paper's reported odor thresholds into an everyday volume of air is ours, done from their published numbers, and flagged as such where it appears.
Heritage ScienceRead 18 minSections 13
Field notePaper Analysis
How to Cook an Egg That Cannot Be Cooked at One Temperature
An egg is a boundary-value problem disguised as breakfast. Its white sets properly around 85°C and its yolk wants to stop cooking near 65°C [1], and a single bath cannot deliver both numbers to a single object at the same time: hold it at 100°C and the white sets while the yolk turns to paste, hold it at 65°C and the yolk is perfect while the white never fully sets, because only one of its proteins, ovotransferrin, is capable of gelling that cool [1]. A 2025 paper in Communications Engineering stops trying to pick one temperature and instead moves the boundary condition itself: the egg spends 2 minutes in a 100°C bath, then 2 minutes in a 30°C bath, repeated 8 times for a total of 32 minutes [1]. Their heat-conduction and reaction-kinetics model finds that the yolk, buried at the centre of the egg, barely notices the swing and settles near a constant 67°C, while the albumen, sitting right against the water, rides the full oscillation and accumulates enough time above its own threshold to set completely [1]. They then cooked real eggs this way and checked the claim with texture analysis, sensory panels, FT-IR spectroscopy and untargeted metabolomics, finding a texture partway between soft-boiled and sous-vide and, unexpectedly, more of several polyphenols retained in the yolk than any conventional method produced [1]. Below we work through the physics ourselves, using nothing but the paper's own reported numbers: why 67°C beats the naive halfway point of 65°C by more than arithmetic averaging would suggest, roughly how far a four-minute thermal cycle reaches into a shell before it fades out, and the fact that the periodic egg spends four more minutes submerged near boiling than a conventional hard-boiled egg does — and still comes out soft.
Heat TransferRead 20 minSections 14
ResearchMethods
Solving Seawater: A Carbonate Equilibrium Solver and What It Says About Acidification
The carbonate system in seawater is four coupled equilibria and one conservation law, and substituting them into each other collapses the whole thing into a single quartic in the hydrogen ion concentration with exactly one positive root. We wrote a solver for it. No seawater was sampled at any point: the computation is the experiment. Using published temperature and salinity dependent constants on the total hydrogen ion scale, the solver satisfies the alkalinity definition to 0.849 ulp of total alkalinity at every one of 64 check points, and a safeguarded Newton iteration, a log-space bisection and the roots of the cleared quartic agree to 1.9 × 10⁻¹⁴ relative. At 18 °C, salinity 35 and alkalinity 2300 µmol/kg, the solver returns pH 8.1817 at 280 ppm and 8.0361 at 420 ppm, against accepted figures of 8.17 and 8.05. Between those two atmospheres the hydrogen ion concentration rises by 39.80%, not the 30% that is usually quoted, and we report that plainly rather than rounding toward it. The gap is 51.5 Monte Carlo standard deviations wide, so the published uncertainty in the constants cannot account for it; asked instead the 370 ppm question the quoted figure was originally answering, our solver returns a pH decline of 0.0993 units against the published observational estimate of 0.11 ± 0.03, agreeing to 0.36 standard deviations. Over the full sweep from 180 to 1000 ppm, bicarbonate rises 30.7% while carbonate falls 69.2% and the aragonite saturation state drops from 4.407 to 1.356. Neglecting the second dissociation, the textbook shortcut, costs up to 0.127 pH units and misstates the aragonite saturation state by as much as 79.5%. A 200,000-trial Monte Carlo over the published uncertainties in six equilibrium constants gives 39.80% ± 0.0004% (standard error), with a 95% interval of 39.43% to 40.17%. Seed 20250320. Runtime 1.33 seconds.
Spring 2025Read 33 minSections 12
ResearchMethods
Where the Textbook Titration Shortcuts Actually Fail
Chemistry courses hand out four or five closed-form shortcuts for titration curves and then guard each one with a shrug. We wrote an exact solver instead, built from charge balance and mass balance with nothing thrown away, and used it to audit the shortcuts across a grid of 57 acid strengths and 41 concentrations. The solver was checked first: charge balance holds to a worst relative residual of 4.35×10⁻¹⁵, speciation fractions sum to one within 1.5 machine epsilons, the strong-acid case reproduces its closed form to 2.6×10⁻¹² pH units, and 60-digit decimal arithmetic reproduces the double-precision answer to 8.9×10⁻¹⁶ pH. 3,489,396 exact solutions were computed in all. Henderson-Hasselbalch then turns out to have a sharply defined licence. At 0.1 M it stays inside 0.05 pH only for acids with pKa between 3.92 and 9.84, and that band slides exactly one pKa unit per decade of dilution in each direction, with fitted slopes of −0.9998 and +0.9995 against a closed-form prediction of −1 and +1. The two edges meet at 1.09×10⁻⁴ M, below which no acid of any strength keeps the buffer equation inside 0.05 pH. Of 18 real acids at 0.1 M, 8 pass and 10 do not; acetic acid is comfortably in at 0.008 pH, formic acid is out at 0.068, phenol is out at 0.065. A Monte Carlo over the region students actually work in, 200,000 draws from seed 20250411, puts the chance of exceeding 0.05 pH at 2.210% ± 0.033%, against 61.77% ± 0.11% over the full parameter space. The five percent rule fails to protect the square-root formula: of the 1,486 grid cells that obey it, 523 still miss by more than 0.05 pH. And the whole analysis is for an ideal solution: switching on Davies activity corrections moves a 0.1 M half-equivalence point by 0.073 pH, which is larger than the tolerance we spent the article defending.
Spring 2025Read 31 minSections 12
ResearchOriginal Research
What a Buffer Is Actually Resisting
Buffer capacity is introduced as a formula and usually abandoned there. We computed it instead, as the exact derivative $\beta = dC_b/d\mathrm{pH}$ of a full charge-balance solver, across dissociation constants from pKa 2 to pKa 12 and concentrations from 0.1 mM to 1 M. No solution was mixed at any point. The club has no laboratory, so the computation is the experiment. Three independent routes to $\beta$ agree: the analytic expression, a five-point finite difference, and a simulated titration that adds a measured dose of base and re-solves the nonlinear equilibrium, which match to 1.4×10⁻¹⁰ and 1.8×10⁻⁸ respectively. The maximum sits at pH equal to pKa, but not exactly: at 0.100 M and pKa 5 it sits 3.477×10⁻⁴ pH units below, a displacement the water term predicts to a ratio of 1.0008, and which is identically zero only at pKa 7.000. Peak height matches $\ln(10)\,C/4$ to 8 parts per million at 0.100 M. Capacity falls to half over a window of 1.531103 pH units, so at one unit from pKa only 33.06% of peak survives and at two units 3.92%. Six of twenty-two tested buffers have no capacity maximum at all. A two-component buffer splits into twin peaks above a separation of 1.143895 pH units, which is $2\,\mathrm{arccosh}(2)/\ln 10$ . For blood at 37 °C, pH 7.40 and pCO₂ 40 mmHg, the model produces 24.50 mmol/L bicarbonate without being told to, and gives an open-system capacity of 57.32 against a closed-system 2.940 mmol/L per pH unit, a ratio of 19.50. A Monte Carlo over five uncertain physiological inputs, 200,000 draws, gives 20.162 ± 0.003 with a standard deviation of 1.213. The open system at 1.3 pH units off its own optimum beats the same closed chemistry sitting exactly on that optimum by a factor of 3.85. Seed 20250321.
Spring 2025Read 27 minSections 12
ResearchMethods
The Common Ion Effect Is Taught Backwards
Every general chemistry course states the common ion effect as a rule with no stated exceptions: add an ion the precipitate already contains, and its solubility falls. We built an exact solver for precipitation equilibria, one that carries the solubility product, stepwise complex formation, the anion's own acid-base speciation, and the water equilibrium as a single coupled system with nothing approximated away, and set it against three sparingly soluble salts. For barium iodate and calcium fluoride the rule holds everywhere on the grid we tried, and the solver's own zero-common-ion limit matches closed-form algebra to a few parts in 10⁷. Silver chloride is where it breaks. Added chloride drives AgCl solubility down to a minimum of 3.11 × 10^-7 M at 3.01 × 10^-3 M added chloride, 223 times below the pure-water value, matching the order of magnitude the analytical-chemistry teaching literature has quoted for this minimum for eighty years [1][2]. Past that point, more chloride does the opposite of what every course says it should: at 5 M NaCl the solver returns 4.96 × 10^-3 M dissolved silver, 15,900 times the minimum and 367 times the pure-water solubility, carried almost entirely as AgCl₂^-, AgCl₃^2- and AgCl₄^3- rather than as free Ag⁺. A hundred thousand Monte Carlo draws over an assumed ±0.05 log-unit uncertainty in the 1952 stepwise formation constants [4] never once erased the reversal: the location of the minimum moved by a few percent, its existence did not. Every mass balance, complex-formation expression and Ksp, evaluated at the converged solution, closes to a few parts in 10¹⁵ (machine precision); an independent full charge balance closes to a few parts in 10⁹, limited by ordinary floating-point cancellation between [H⁺] and [OH^-] near neutral pH, not by the model. Calcium fluoride shows a second, unrelated failure of the naive picture: at fixed added fluoride, dropping the pH from 7 to 0 raises its solubility 816-fold through protonation of F^-, a channel the common-ion rule never mentions and one that, for this salt, never reverses. The textbook rule is not false, only incomplete: it is missing a clause. Seed 20250306.
Spring 2025Read 27 minSections 12
ResearchOriginal Research
The Reaction That Gives You the Wrong Product Because You Were Impatient
Two products, one shared starting material, two reversible one-step reactions, and no way for the products to talk to each other directly: this is the standard textbook picture of kinetic versus thermodynamic control, and the textbook almost never tells you how long the experiment takes. We wrote the rate equations, integrated them with a from-scratch fourth-order Runge-Kutta stepper, and separately solved the same linear system exactly by eigendecomposition, then checked the two against each other and against the closed-form single-reaction solution at matching time points. They agree to better than one part in a million everywhere we tested, RK4's global error falls with step size at an observed order of 4.125 against a theoretical 4.000, mass is conserved to 2.0 × 10⁻¹⁴ relative, and the simulated ratio of products reproduces both textbook limits: the kinetic ratio $k_2/k_1$ at short times and the thermodynamic ratio $K_2/K_1$ at long times, each to better than one part in ten thousand once the numerics stop fighting catastrophic cancellation. For a representative pair of channels, 10 kJ/mol apart in barrier height and 25 kJ/mol apart in stability, the faster product claims 98.0% of the mixture within two hours, then loses ground it never gets back: the slower, more stable product does not overtake until 111.7 days at 298.15 K. Across a grid of 676 barrier and stability combinations, only 3.0% cross within a day and 22.5% never cross at all because we gave the "thermodynamic" product no real stability edge; the rest take anywhere from two weeks to millions of years, with almost nothing in between a day and two weeks, because the crossover time is exponential in the barrier gap and exponentials do not fill in the middle. Heating the baseline reaction from 298 K to 400 K divides the wait by 137,000; cooling it to 250 K multiplies the wait by 7,620. No reagent was weighed and no flask was heated for this study. The computation is the experiment, and the club has tried to be honest about what it leaves out.
Spring 2025Read 25 minSections 12
Field noteOriginal Research
Why the Periodic Table Is That Peculiar Shape
Two independent constructions, neither of which involves a beaker. The first builds the periodic table out of nothing but quantum numbers: enumerate every orbital with $l<n$ , give each one $2(2l+1)$ seats, and fill them in order of increasing $n+l$ . Run that rule from hydrogen onward and it reproduces the period lengths 2, 8, 8, 18, 18, 32, 32 exactly and lands on all seven noble-gas atomic numbers — 2, 10, 18, 36, 54, 86, 118 — with no chemistry supplied as input anywhere. Laid out honestly, with the f-block left in line rather than cut out, the table is 32 columns wide; the 18-column chart on the classroom wall is that same table with 14 columns lifted out and filed underneath. Checked element by element against measured ground states from Z = 1 to 103, the rule gets 83 of 103 right and misses 20, and every miss sits in the d-block or the f-block — the s- and p-blocks, 44 elements between them, never break it once. The second construction asks how far a much cruder rule, Slater's 1930 electron-screening recipe, can push those same quantum numbers toward numbers you could look up: effective nuclear charge, ionisation energy, atomic radius. It passes the one check with an exact answer — hydrogen predicted at 13.6057 eV against a measured 13.5984 eV — and then degrades on a schedule: a mean ionisation-energy error of 11.9% across period 2 widening to 38.4% across period 3, worst at aluminium (+79.4%), because the rule has no way to ask which orbital an electron actually occupies. The matching radius law does better, landing within 11.2 pm on average of Slater's own 1964 empirical radii from one fitted constant. A last, even cruder relativistic estimate on a single 1s electron is enough to see, in outline, why a gold atom absorbs blue light and a silver one does not.
Chemistry & PhysicsRead 24 minSections 12
Field notePaper Analysis
The Core Is Leaking: Ruthenium-100, Hawaiian Lava, and Gold From the Center of the Earth
Earth's metallic core is supposed to be a sealed room: whatever fell into it 4.5 billion years ago stays there. In May 2025 a Göttingen-led team reported that Hawaiian basalts carry an excess of the isotope ruthenium-100, about nine parts per million, a signature that matches core metal and matches nothing else on the planet. The same rocks carry the tungsten-182 deficit you would expect from that source, an independent clock nobody tuned to match [1]. We re-derive the mixing arithmetic ourselves in a deliberately simple two-component model and get a source containing 0.071% core metal by mass (68% interval 0.044–0.129%), comfortably inside the paper's published upper bound. A mass-flux calculation turns that into roughly 159 kg of core-derived gold a year. One wedding band's worth every thirteen minutes. What follows is a student reconstruction rather than the paper's own analysis, built on seven stated assumptions, one of which (the isotopic composition of a reservoir nobody has ever sampled) could move the answer by a factor of three.
GeochemistryRead 32 minSections 12
Field noteExplainer
Younger Rock Beneath Older Rock: The North Sea Layers That Sank Through Time
A 2025 study of basin-scale 3D seismic data from the northern North Sea reports hundreds of kilometre-wide sand bodies that appear to have liquefied, sunk through a 200–500 m interval of older microfossil ooze, and come to rest beneath it. Rafts of that older ooze went up instead. They built mounds. The authors call the sunken sand sinkites, the risen ooze floatites. Grant the interpretation and the law of superposition has its largest documented breach. We rebuilt the physics with a deliberately simple club model, which puts saturated sand at 40% porosity at 2002 kg/m³ and ooze at 65% porosity at 1404 kg/m³, a contrast of 598 kg/m³ pushing steadily downward at 5861 N/m³. Liquefaction wants a pore-pressure ratio of 1.00 and nothing else. A Stokes-type descent of 200 m takes 2 kyr to 2 Myr for effective viscosities of 10¹⁷–10²⁰ Pa s, and nobody has ever measured that viscosity. The mechanism is allowed. What happened is a different question, still open.
GeologyRead 26 minSections 11
Field notePaper Analysis
Caught on CCTV: The First Video of a Fault Tearing the Ground Apart
On 28 March 2025 a security camera twenty metres from Myanmar's Sagaing Fault filmed the ground tearing open. One side slid past the other. Frame-by-frame pixel tracking gives 2.5 ± 0.5 m of slip in 1.3 ± 0.2 s, peaking at 3.2 ± 1.0 m/s. It throws in something nobody ordered. The slip path curves [1]. Nobody had ever filmed a fault rupturing. We rebuild the pulse from the three published numbers. They will not all fit the textbook crack-tip slip function at once. The implied shape factor comes out at k = 0.601. Every regularised Yoffe pulse sits at k ≤ 0.5. Our own smoothed-ramp model hits all three exactly. It puts peak slip acceleration at 1.29 g. Shear strain rate across the mole track: 2.1 per second, roughly 10^14.6 times the tectonic loading rate. A rotating slip vector fitted to the paper's rake data bows 19.5 cm off the straight line. At this camera's scale, 2.58 pixels. And nearly all of the turning finishes before peak velocity. Everything past §3 is the club's own simplified forward model rather than the authors' analysis, and we say so every time it matters.
SeismologyRead 35 minSections 12
ResearchReplication
Measuring a Universal Constant on a Laptop: The Site Percolation Threshold
Fill the cells of a square grid at random. Each cell is occupied with probability $p$ , independently. Below a certain density the cells form scattered islands. Above it, one cluster reaches from top edge to bottom. That density, on an infinite grid, is the site percolation threshold. The square lattice has no closed form for it, so the constant has to be measured. We measured it. Our experiment is a computation, and nothing was observed in the physical world: a club program ran 114,520 Newman-Ziff union-find sweeps on lattices from $L = 16$ to $L = 1024$ , in 526.5 seconds of wall clock, seeded from 20251215. Extrapolate the half-crossing point $p^*(L)$ to infinite size with the exact two-dimensional exponent $1/\nu = 3/4$ , and the intercept comes out at $p_c = 0.5928171 \pm 0.0001110$ , the error bar bootstrapped from the sweeps themselves. The accepted value is 0.59274605. We sit +0.0000710 above it. Call that +0.64 of our own error bars. The disagreement first appears in the fifth decimal place, and three checks stand behind the number. The union-find matched a separate flood fill on 11,400 random configurations, with zero mismatches, and the pipeline reproduced the exact crossing polynomial of the 3×3 and 4×4 lattices to within 0.5 standard errors at every point tested. And the crossing probability at the accepted threshold settled onto a size-independent 1/2, as conformal field theory says it must, moving from 6.60 standard errors high at $L = 16$ to 0.08 at $L = 256$ . The fractal dimension of the largest cluster came out at 1.8823 against the exact 91/48 = 1.8958, a shortfall we keep and explain rather than fit away. Including our smallest lattice moves the answer to 0.5930250, 2.92 error bars out, and that single modelling choice is the largest one in the study.
Winter 2025Read 25 minSections 12
ResearchOriginal Research
The Same Disease on Four Different Networks
Give four populations the same disease and the same average number of contacts each, then change nothing but the shape of the contact network. Nobody was infected, contacted or observed. The computation is the experiment. It ran on a school laptop, and a rerun finishes in minutes. We simulated stochastic SIR epidemics on four graphs of 10,000 nodes with mean degree fixed at 6, a triangular lattice, an Erdős-Rényi random graph, a Watts-Strogatz small world and a Barabási-Albert scale-free network, at 600 runs per configuration across fourteen transmission rates, plus refinement, validation and convergence sweeps, 80,700 simulated epidemics in all. At a matched transmissibility of T = 0.50 the four networks agreed on the final attack fraction to within a factor of 1.06 and disagreed about everything else: peak prevalence ranged from 0.0602 on the lattice to 0.4406 on the scale-free network, a factor of 7.32, and time to peak ran from 12.8 steps to 136.6, a factor of 10.69. Measured epidemic thresholds spanned a factor of 5.14, from T_c = 0.0728 on the scale-free network to 0.3741 on the small world, against a well-mixed prediction of 0.1667 for all four. The random graph landed on the Molloy-Reed value 0.1665 to within 0.0014; the lattice and the small world sat 1.76 and 1.89 times above it. Validation: the mass-action limit reproduced the Kermack-McKendrick final size equation to within 0.0074 at every R₀ above 1.25 and to 0.0011 above 1.5, and a constant-infectious-period lattice recovered the exactly known triangular bond percolation threshold 2 sin(π/18) = 0.34730 as 0.33068 ± 0.00875, 1.90 standard errors low, with finite-size scaling across L = 40 to 140 showing the gap closing as L^−3/4 exactly as percolation requires. Seed 20251215.
Winter 2025Read 28 minSections 12
ResearchOriginal Research
When Most People Infect Nobody: What Overdispersion Does to an Outbreak
An outbreak's average reproduction number, $R$ , does not say how that transmission is shared out. We fixed $R = 2.5$ and swept the negative binomial dispersion parameter $k$ across seventeen values from 0.01 (extreme superspreading) to 10,000 (indistinguishable from Poisson, no superspreading at all), running 40,000 independent branching-process replicates per value and checking every simulated extinction probability against the exact analytic root of the offspring distribution's probability-generating function. They agreed everywhere: the largest deviation across all seventeen cells was 2.26 standard errors, and none exceeded 3. At the realistic value $k = 0.1$ , reported for SARS and SARS-CoV-2, a lone introduction died out on its own 86.11% ± 0.17 percentage points of the time. At $k = 10{,}000$ , the near-homogeneous case with the identical average, it died out only 10.51% ± 0.15 points of the time, matching the classical Poisson branching-process result (0.1074) to 3.65×10^-5. Same average transmission, nine-fold difference in how often the chain simply stops. Among the cases that transmit at all when $k = 0.1$ , the most active 10% account for 81.9% of all onward infection. We then compared two ways of spending an identical unit of control effort, defined as an equal fractional cut to the uncontrolled mean: thinning every case's transmission uniformly, against capping only the largest gathering-driven events to hit that same mean. Across thirteen effort levels, uniform thinning produced a higher extinction probability than capping at eleven of them, by up to 5.2 percentage points ( $z = 19.4$ ) at 45% effort, reversing only at the very edge of guaranteed elimination. No person, gathering or pathogen was involved in any of it; every number above is the printed output of a program the club wrote and ran on a laptop. Master seed 20251217.
Winter 2025Read 24 minSections 11
Field notePaper Analysis
Solving a Ten-Year Marine Murder Case: What Actually Killed Five Billion Sea Stars
Sea star wasting disease began in 2013. It emptied the Pacific coast of North America of sunflower sea stars. Roughly 3,000 km of coast. The prime suspect for ten years was a virus. Wrong suspect. In August 2025 a team led by Melanie Prentice named the killer: Vibrio pectenicida strain FHCF-3, cultured from the coelomic fluid of a diseased star, able to reproduce the disease in healthy animals and recoverable from them afterwards [1]. The evidence is a plain walk through Koch's postulates, done in flowing seawater. This field note reconstructs the case. It follows the filtration step that quietly killed the virus hypothesis, and then goes further than the paper does. We built a deliberately simple three-species model of a kelp reef, our arithmetic and not theirs, and it says that a 96.2% loss of sea stars flips the modelled forest into an urchin barren within 2.7 years. Putting the kelp back inside forty years costs more still. 55% more sea stars than were there before the outbreak. A second club model walks an epidemic down a coastline at 8.22 km per day. It returns a reproduction number of 1.57. That number is too small to explain the mortality anyone observed. We take the mismatch seriously. The strongest published objection to the bacterial result gets its own section.
Disease EcologyRead 34 minSections 12
ResearchOriginal Research
Why Small Fires Are Common and Big Fires Are Not Rare Enough
Trees grow at random, lightning falls at random, and whatever connected patch the bolt lands in burns. Those two rules make the Drossel-Schwabl forest fire model, which the literature quotes everywhere as a clean case of self-organised criticality, complete with a power-law distribution of fire sizes. We tested the claim rather than repeating it. The club wrote the model and ran it on a laptop, and since no forest was observed and no fire measured anywhere, the computation itself is the experiment. Across five values of $\theta = p/f$ spanning two decades and three lattice sizes we recorded 416,000 fires, 120,000 of them in a main run at $512 \times 512$ with $\theta = 500$ . We fitted a truncated discrete power law by maximum likelihood, after first checking the estimator on synthetic data with a known exponent, where it recovered the truth in all five tests with a largest $z$ of 1.61. The same synthetic data fitted the popular way, a straight line through a log-log histogram, came out 35.2% low on a known $\tau = 1.5$ , which is the sort of error nobody notices when the true answer is unknown. On the fire data, the main-run exponent over $s \in [4, 500]$ is $\tau = 1.1269 \pm 0.0025$ (moving-block bootstrap), close to the $1.15$ the pre-2002 literature reports. The power law is nonetheless rejected: the Kolmogorov-Smirnov distance of 0.00824 gives $p < 0.005$ over that window, and over the full range the distance is 26× the 95th percentile of what the fitted law itself produces. The exponent is not a constant of the model either, which is the part that unsettled us. It climbs from 1.0394 at $\theta = 50$ to 1.1741 at $\theta = 5000$ , a drift of +0.069 per decade, with $\chi^2 = 466.3$ on 4 degrees of freedom against a constant. Two things did land: the steady-state identity $\langle s \rangle = \theta(1-\rho)/\rho$ held to 0.14%, and our tree density at $\theta = 125$ , 0.38073, sits 0.9 standard errors from Grassberger's tabulated 0.379837. Our lattices are three to four orders of magnitude smaller in area than the ones that reopened the criticality argument, so we can measure the drift and cannot settle where it stops. Seed 20251124.
Winter 2025Read 32 minSections 12
ResearchOriginal Research
Can You Tell a System Is About to Tip Before It Does
A system approaching a fold bifurcation recovers from disturbance more and more slowly, and that slowing is supposed to be readable in ordinary noisy measurements before anything visible happens. The club has no laboratory, so we tested the claim by simulation, which is in any case the only place where the control can be run. In 38,600 simulated trajectories of a cusp normal form driven slowly toward its fold under process and observation noise, the signal is unmistakably there: mean lag-1 autocorrelation climbs from 0.331 to 0.691 and variance triples, both far outside their standard errors, and both match the analytic Ornstein-Uhlenbeck prediction once the known bias of a 100-sample detrended window (0.8722, measured separately on a stationary control) is divided out, with a corrected ratio of 1.002 at the start of the run. The measured recovery rate lands 0.24% from $\kappa = 3x^{*2}-1$ and 0.76% from the exact value for the Euler scheme we actually integrated, which brackets the fit between the two predictions. Usefulness is another matter. A textbook one-sided Mann-Kendall trend test at nominal $\alpha = 0.05$ , run on overlapping sliding windows the way it usually is, fires on 30.17% ± 0.84% of controls that are not approaching anything, 30 standard errors from nominal. Removing the window overlap restores the nominal rate (5.20% ± 0.41%). The remaining problem does not go away. A control driven at the same rate that stops 10% short of the fold triggers a calibrated 5% detector on 20.53% of runs against 21.07% power on runs that genuinely tip, an AUC of 0.502. Across ten configurations spanning noise, driving speed, window length and observation noise, the mean gap between detection power and the near-miss alarm rate is 0.0127. The indicator measures how close a system is to a threshold. It says nothing about whether the system will cross it. Seed 20251217.
Winter 2025Read 33 minSections 12
Field notePaper Analysis
The Iron Ball at the Center of the Earth Keeps Changing Its Skin
Nobody has ever touched Earth's inner core, and nobody ever will. Every fact we own about it arrives secondhand, carried the last 5,150 kilometres by a seismic wave. In February 2025, a team led by John Vidale re-examined a dataset built for a different purpose — 121 repeating earthquakes near Antarctica's South Sandwich Islands, recorded between 1991 and 2023 by two seismic arrays in the far north — and asked a narrower question than usual: when the inner core's own drifting rotation brings it back to a position it already occupied years earlier, do the two recordings match [1]? At one array, Yellowknife (YKA), whose waves graze the inner core's outer few hundred kilometres, some pairs recorded between 2004 and 2008 did not match. At the other, Eielson (ILAR), whose waves bottom out deep near the centre, the same pairs matched fine [1]. The clean explanation is that whatever changed did not change everywhere — it changed near the boundary, on a timescale of a few years, while the deep interior stayed put. We rebuild the geometry that makes this comparison possible, work out why a solid iron sphere is not supposed to do this, and then do the club's own arithmetic: the inner core is about 35% the volume of the Moon, and the thin, structurally softer "mushy" layer that a separate 2017 study found elsewhere on the same boundary is only about 0.3–0.7% of the inner core's radius — a few kilometres of restless skin wrapped around twelve hundred kilometres of otherwise very patient iron.
GeophysicsRead 23 minSections 13
Field notePaper Analysis
The Sandwich Compound That Raced Its Own Half-Life
In February 2025 a team working across Lawrence Berkeley National Laboratory and Oak Ridge reported the first structurally confirmed carbon-berkelium bond: a metallocene, an ion clamped between two flat carbon rings like the meat in a sandwich, built around the intensely radioactive, synthetic element berkelium [1]. What makes this remarkable is less the molecule than the material it is made of. The whole synthesis ran on 0.3 milligrams of berkelium-249, an isotope with a half-life of roughly 327 days [2] that the world produces in batches of about 13 milligrams every couple of years [9], and by the researchers' own account, the 0.3 mg they had on hand already came mixed with 0.2 mg of californium-249 — the isotope's own decay product — before the reaction even started [9]. This note is about three things: what a "sandwich compound" structurally is, why doing wet chemistry on a sample that is actively transmuting into a different element counts as an extraordinary experimental feat, and what it means that the resulting berkelium-carbon bond turned out to hold onto real covalent character deep into a part of the periodic table where chemists mostly expected ionic, lanthanide-like behaviour. Along the way we run our own numbers: from the reported 0.3 mg/0.2 mg berkelium/californium mixture and the published half-life, we back out roughly how long the isotope had been sitting since it was last purified, and we hold 0.3 milligrams up against a few things you can picture without a periodic table.
Actinide ChemistryRead 19 minSections 12
Field noteExplainer
How to Read a Scientific Paper Without Pretending You Understood It
Reading primary research is the one skill this club exists to teach. Almost nobody gets taught it. Here is the handout we wish somebody had pushed across the table on day one: what each part of a paper does, why experienced readers go to the figures before the prose, how to check whether the statistics carry the conclusion, and what peer review promises in writing. Honesty demanded that we measure the thing we were describing. With a text-analysis script we wrote ourselves, we scored 67 open-access research papers pulled from Europe PMC and all 298 sections of the club's own 27 published articles on one instrument. The median Results section runs 3,148 words against a median abstract of 206, a compression of 15.3 to one. Register changes harder than length does. Results sections favour reporting vocabulary over interpretation vocabulary 7.58 to one; abstracts flip it, with interpretation winning at 0.75 to one, a swing of about ten times. The popular charge that abstracts inflate their verbs did not survive our own test. Paired within each paper, abstract-minus-Results claim strength came out at +0.33 per thousand words, t = 0.16, n = 67, which is nothing at all. Abstracts do not shout. They change what kind of sentence they are writing, and they take the evidence away with them. Read for that.
Scientific PracticeRead 37 minSections 13
ResearchOriginal Research
How Often Does a Good Gene Lose? Measuring the Reach of Genetic Drift
A mutation that makes its carrier one percent fitter still disappears almost every time it appears. We measured how often. The instrument was a Wright-Fisher simulation the club wrote and ran on a laptop, and no organism entered the study at any point, because here the computation is the experiment. Across a grid of eight selection coefficients ( $s = 0$ to $0.1$ ) and five population sizes ( $N = 50$ to $10{,}000$ ) we ran 11,932,191 independent replicates from a single new mutant copy and counted 90,675 fixations, holding the relative standard error to no worse than 5.93% in any of the 40 cells. For $N = 1000$ and $s = 0.01$ , a single new copy fixed in 1.992% ± 0.063% of replicates. Loss took the other 98.01%. The neutral case recovered the exact answer $1/(2N)$ at all five population sizes, the one corner of the study where the theory leaves no room for approximation at all, and every 95% interval covered the true value, the largest z score being 1.33. Haldane's $2s$ is excellent while $4Ns \gtrsim 4$ . Below that it fails badly, falling short of the measured value by 488% at $N = 50, s = 0.001$ ; its mean absolute error is 255.6% for $4Ns < 1$ and 2.5% for $4Ns$ between 4 and 20. Kimura's diffusion formula holds to a few percent across the whole grid, then breaks at the opposite edge: at $s = 0.1$ it sits 3.0% above our measurement, a deviation confirmed independently at all five population sizes with a pooled z of −7.33. Haldane's own branching process, taken before the small- $s$ step that produces $2s$ and solved numerically instead, matches that same measurement to 0.2%, which puts the fault in the diffusion limit rather than in either man's reasoning. Conditional fixation times reproduced the $4N$ neutral prediction to within 5.1% at every size. Seed 20240917.
Fall 2024Read 31 minSections 12
ResearchOriginal Research
A Population That Grows on Average Can Still Die Out
A population in which every individual leaves 1.15 offspring on average grows by 15 per cent a generation and can still go extinct by bad luck alone. We simulated a Galton-Watson branching process and a continuous-time birth-death process, 20,000 replicates per parameter cell, for starting sizes from 2 to 500 and five offspring distributions whose variance ranges from 1.15 to 33.2 with the mean held exactly fixed. The probability of extinction within 100 generations falls below 5 per cent at a starting size of 10.5 under Poisson reproduction, 21.4 under geometric reproduction, and 302.9 under a lottery distribution in which about one individual in twenty-six leaves 30 offspring and the rest leave none. Across the five levels the threshold is close to proportional to the variance: a log-log fit gives a slope of 1.0014 with R² = 0.9994, so the safe size is about 8.8 times the offspring variance at this growth rate. Every simulated cell carries an exact analytic partner, computed along a wholly separate path: we iterate the offspring probability generating function forward, then locate its smallest non-negative fixed point by bisection, and the two answers have to agree. Over the 154 cells with at least ten expected extinctions the mean z-score is −0.048 with a standard deviation of 0.943 and no cell exceeds three standard errors; the two closed forms we can solve independently, 1/m for geometric offspring and the Lambert-W form for Poisson offspring, match our root finder to within 1.3×10⁻¹⁵. Haldane's 1927 approximation for lineage survival, 2s/V, written before anyone could simulate anything, sits within 0.2 to 13.2 per cent of our exact values over a thirty-fold range of variance. The whole study is a computation: the club has no laboratory, observed no animals, and claims nothing about any of them.
Fall 2024Read 28 minSections 12
ResearchOriginal Research
Everything a Small Founding Group Loses on the Way
A founding group of two individuals carries 74.973% ± 0.048% of the source population's heterozygosity and 20.69% of its alleles. Both numbers come out of one run of one program. They are most of what this paper has to say. We simulated founder events by drawing $2k$ gametes from a source population with a fixed and printed allele frequency spectrum, then grew the new population back to a carrying capacity of 1,000 under Wright-Fisher reproduction with K-allele mutation. No organism was involved, and no island. The computation is the experiment. Across fifteen founding sizes from $k = 2$ to $k = 500$ at 20,000 replicates each, measured heterozygosity retention matched the closed form $H_{\text{new}} = H_{\text{old}}(1 - 1/2k)$ with a largest standardised deviation of 2.13 and no cell outside three standard errors. Allele loss matched the binomial prediction $\sum_i (1-p_i)^{2k}$ to a largest deviation of 1.54 standard errors; two founders lost 183.21 of the source's 231 alleles against a prediction of 183.19. Drawing founders without replacement from a finite source of $N = 60$ reproduced the hypergeometric correction, and at $k = N$ the founding group lost nothing at all, in every one of 20,000 replicates, to a floating-point difference of $1.11 \times 10^{-16}$ . Over a grid of eight founding sizes by four growth rates, 5,000 replicates each, simulated gene diversity agreed with an exact recursion at every one of the 32 cells, mean $z$ −0.004 and standard deviation 0.998. Growth rate mattered more than we expected: two founders growing at 5% a generation held 12.35% of the source's diversity by the time they reached capacity, and at 100% a generation they held 60.15%. The bottleneck signature faded much faster than the loss did. A heterozygosity excess test on 30 individuals at 15 loci detected a five-founder event 67.08% of the time at generation 10 and 6.48% of the time at generation 60, while the population itself was still 29.6% short of the control two hundred generations later. Seed 20241108, 630,000 replicates, 276 seconds.
Fall 2024Read 29 minSections 12
ResearchMethods
Absence of Evidence: How Many Surveys Before You Can Say a Species Is Gone
A survey that finds nothing has not proved anything. We built the detection half of a single-season occupancy model and asked the scheduling question underneath it: how long does a run of blank visits have to be before finding nothing counts as evidence of absence? With a constant per-visit detection probability $p$ , the probability that an occupied site produces $k$ blanks is $(1-p)^k$ , so 95 percent confidence needs 14 visits at $p = 0.20$ and 59 at $p = 0.05$ . We checked 500,000 simulated survey histories per configuration against that closed form across 76 cells: mean $z = +0.0044$ , largest $|z| = 3.32$ , and the eleven independent anchor cells average $+0.27$ , which is 0.89 standard errors from zero. Varying detection probability costs nothing, or costs everything, depending on where the variation lives. Redrawn from a Beta distribution at every visit it is free, which we prove and then confirm. Drawn once per site and kept, the bill arrives: at mean $p = 0.20$ and Beta concentration $\kappa = 5$ (SD 0.163) the requirement rises from 14 visits to 76, a factor of 5.66, and at $\kappa = 2$ to 2,369 visits. Twenty blanks that the naive formula scores at a 1.15 percent chance of error actually leave 16.7 percent, and we validated that heterogeneous case three ways, simulation against a Beta-function closed form against tanh-sinh quadrature, agreeing to 5.7 × 10^-14. Our first quadrature was wrong by 0.86. We say where. Putting Kéry's published snake detection probabilities into our formula reproduces two of his three survey requirements exactly. Seed 20240921. Everything here is computation, and the club made no observations of anything.
Fall 2024Read 27 minSections 12
ResearchMethods
How Many Samples Before Your Species Count Means Anything
A species count is always short. The four estimators everybody reaches for are supposed to tell you by how much, so we tested them against a truth we knew, because we built it: five community types, each holding exactly 100 species, sampled at ten levels of effort, 6,000 replicate communities per configuration, 300,000 simulated surveys in total. No organism was involved at any stage. On a mildly uneven community (log-normal $\sigma = 1.0$ , Pielou $J = 0.895$ ) a survey of 500 individuals turned up 86.28 ± 0.05 species, which Chao1 corrected to 95.42 ± 0.09. Still 4.58 species short. Chao2 landed at 94.74, ACE at 95.49. The second-order jackknife, a cruder and older device, overshot to 104.26 and reached the 10 percent band at 106 individuals against Chao1's 268 and Chao2's 302, a gap of nearly three to one. Chao1 and Chao2 run low in all fifty cells of the grid, the jackknives correct past the truth instead, and the ranking tracks the abundance distribution rather than the estimator's reputation. Past a certain effort the correction stops helping: at n = 5,000 in that same community the raw count beat Chao1 on root mean square error, 0.490 against 0.570. Validation: the accumulation curve matched its closed form in all 50 cells, largest $|z| = 2.62$ , spread of $z$ across cells 1.01; Hurlbert's hypergeometric rarefaction matched direct subsampling to 0.71 standard errors at worst; and every estimator returned exactly 100.000 species at exhaustive effort. Seed 20241104.
Fall 2024Read 27 minSections 12
Field noteMeta-analysis
The Insect Decline Literature Disagrees With Itself, and We Wanted to Know How Badly
One German study reported a 76.7 percent fall in flying-insect biomass over 27 years [1]. A synthesis of 166 long-term surveys put terrestrial declines near 9 percent per decade while freshwater insects rose about 11 percent per decade, and an erratum later steepened the terrestrial figure [2, 3]. More than 5,300 American time series gave net trends indistinguishable from zero [6], which is a statement about time series and about the units they were expressed in, though almost nobody repeated it that way. All three are competent work. We extracted 11 effect sizes from 6 papers covering 10 monitoring datasets, traced every one to a printed slope or a printed interval in its source, converted them to a common log-linear annual rate, and pooled them under a DerSimonian-Laird random-effects model. The pool comes out at −2.68 percent per year (95% CI −4.34 to −0.99), or −23.8 percent per decade, and the spread around it is the part worth printing: I² = 96.3 percent, with Cochran's Q at 269.96 on 10 degrees of freedom. That second pair of numbers matters more than the first. A 95% prediction interval for the next comparable study runs from −8.14 to +3.10 percent per year, so a new paper reporting insects on the rise would sit comfortably inside everything published to date. We also simulate the objection that monitoring programmes began at unusually good sites, where regression to the mean manufactures a decline nobody caused; under a deliberately unkind version of that model the artefact accounts for roughly 15 percent of our pooled estimate. Real and quantified, and nowhere near large enough to carry the objection. We are a school club reading printed tables. Weight the pooled number below what its confidence interval implies. Weight the heterogeneity above it.
EcologyRead 41 minSections 13