Are the best discovery candidates actually true?
The 5 best candidates of sets 1 and 2 (items in the top 10 under every judge panel first), and the 10 best of set 3 (philosophy of science, added 20 Oct), were checked independently by three strong models, each asked to restate the claim precisely, re-derive it from scratch rather than trust the text, hunt for counterexamples, attempt a full proof or disproof, and judge novelty and fairness. For two items I added my own check.
60 reviews: 32 "correct", 28 "correct with caveats", 0 other. Set 3: 30 reviews, 13 "correct", 17 "correct with caveats", none incorrect, but every verifier called every set-3 item a special case or variant of known results (one "known"): sound, not new. Caveats: all items were written by GPT-6.1, and GPT-6.1 is one of the three verifiers; "new as far as I know" is the models' recollection, not a literature search; nothing here is peer review.
Adding one passive spring can increase both normal-mode frequencies and reduce a sensor’s equilibrium displacement noise by exactly one third, yet amplify its response to a spatially uniform static force without bound.
supporting text
Consider two equal-mass mechanical coordinates x and y, with units chosen so their original potential is U₀=(x²+a²y²)/2, where 0<a<1. Apply the same static force f to both coordinates. Initially, the mean displacement of x is f, and its thermal variance is kBT. Now add an ordinary positive spring through an ideal massless lever, contributing (x−ay)²/2. The new stiffness matrix is K=[[2,−a],[−a,2a²]], whose inverse is [[2/3,1/(3a)],[1/(3a),2/(3a²)]]. Consequently, the variance of x becomes exactly 2kBT/3, while its mean displacement becomes f(2+1/a)/3. The response amplification diverges as a approaches zero. The added stiffness is positive semidefinite. Moreover, its direction overlaps both original normal modes, so both squared normal-mode frequencies strictly increase for every 0<a<1. Thus neither a softened mode nor active feedback explains the amplification. There is no conflict with fluctuation–dissipation: the reduced variance measures the response to a force applied to x alone. A uniform force also acts on y, and the lever transfers its large displacement into x. This construction separates three quantities often conflated in sensor design: local thermal noise, modal stiffness, and sensitivity to a distributed load.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | All computations (K⁻¹, 2/3 variance, f(2+1/a)/3 response, both eigenvalues strictly rising) check out, but the unbounded amplification relies on a soft original mode that stays soft and is merely lever-coupled into x, so the effect is a pedagogical instance of interlacing plus lever amplification rather than a new or general sensor principle. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: partly | The algebra, mode-stiffening, and FDT reasoning all check out, but 'one third' refers to variance (not rms), the divergence holds only over the parameter family as a→0, and the construction needs an ideal lever and an equal-force convention. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: partly | The idealized construction is mathematically correct, but the one-third reduction concerns variance, and unbounded uniform-load response requires a singular family with an increasingly soft auxiliary coordinate. full review |
For a stationary reversible finite-state hidden Markov process with real-valued deterministic outputs, matching the independent-output distributions of just the two-term and three-term consecutive sums forces the entire output sequence to be independent.
supporting text
Let Z_t be a stationary reversible finite-state Markov chain, with positive stationary probabilities, and let X_t = f(Z_t). Suppose X_0+X_1 has the distribution of the sum of two independent copies of X_0, and X_0+X_1+X_2 has the corresponding three-copy distribution. No individual transitions or hidden states need to be observed. For a real parameter t, put m = E[exp(tX_0)] and W(z) = E[exp(tX_1) | Z_0=z]. Reversibility and the Markov property imply that, conditional on the middle hidden state, its two neighbors are independent and have the same conditional distribution. Consequently, E[exp(t(X_0+X_1))] = E[exp(tf(Z_0))W(Z_0)], E[exp(t(X_0+X_1+X_2))] = E[exp(tf(Z_0))W(Z_0)²]. The assumed sum laws make these expressions m² and m³. Expanding a square therefore gives E[exp(tf(Z_0))(W(Z_0)−m)²] = m³−2m³+m³ = 0. Its weight is strictly positive, so W(z)=m at every state. This holds for every real t. Since the output has finite support, equality of these exponential transforms means that the next output's conditional distribution is identical at every hidden state. The Markov property now makes each next output independent of the complete output past. Induction proves that the whole observed sequence is independent and identically distributed. Reversibility is the crucial mechanism turning aggregate sum information into a nonnegative square.
| claude-fable-5.1 | correct | new as far as I know | fair: yes | The argument is valid: reversibility makes E[e^{t(X_0+X_1+X_2)}] = E[e^{tf(Z_0)}W(Z_0)^2], the two sum-law hypotheses force E[e^{tf}(W−m)^2]=0 hence W≡m for all t, so the next output is independent of the hidden state and the sequence is iid. full review |
| claude-opus-5.5 | correct | new as far as I know | fair: yes | The reversibility-based perfect-square argument is valid: matching the 2-sum and 3-sum laws forces E[e^{tX_1}|Z_0=z] to be constant in z for all t, which makes the observed sequence i.i.d. full review |
| gpt-6.1-sol | correct | new as far as I know | fair: yes | The two stated sum laws do force iid outputs under the given reversible finite-state assumptions; the proof is sound, but historical novelty is unverified. full review |
My own check. I checked the proof line by line. Conditioning on the middle hidden state, reversibility makes its two neighbours i.i.d. with law P(z,·), so E e^{t(X₀+X₁+X₂)} = Σ π(z)e^{tf(z)}W(z)², while E e^{t(X₀+X₁)} = Σ π(z)e^{tf(z)}W(z). The two hypotheses give m² and m³, so Σ π e^{tf}(W − m)² = 0, hence W ≡ m for all t: the next output's law does not depend on the hidden state, and independence follows by induction. Correct. Novelty: none of the three models knows a published statement, but no literature search was done.
Costless, complete income equalization can create an attracting permanent boom–bust cycle and cut average total income by one third in a two-household economy that otherwise reaches its unique steady state within two periods.
supporting text
Consider a reduced-form economy with two households. Household B earns 1 each period. Household A’s next-period earnings depend on B’s current disposable income y through f(y)=5 exp[−k(y−1)], where k=(ln 5)/2. This decreasing relationship represents a lagged competitive externality: greater resources for B reduce the earnings opportunity available to A. Earnings are positive and smooth, and the production rule does not change with policy. Without redistribution, B’s disposable income becomes 1 immediately, so A’s subsequent earnings become f(1)=5. From any starting point, the economy reaches the unique income vector (5,1) within two updates, producing total income 6 thereafter. Now pool and split all earnings equally after every update, without administrative costs. Both households have disposable income y, whose dynamics become y_next=[f(y)+1]/2. Since f(1)=5 and f(3)=1, disposable incomes alternate exactly between (1,1) and (3,3). This cycle is locally attracting: the product of the two update derivatives is 5(ln 5)²/16, approximately 0.809, below one. Total income therefore alternates between 2 and 6, averaging 4 rather than 6, while measured inequality is always zero. Redistribution closes a feedback loop that did not previously exist. The construction does not claim that equalization generally destabilizes economies; it isolates a smooth, explicitly checkable lagged-externality mechanism.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | The arithmetic and stability analysis are correct (f(1)=5, f(3)=1, multiplier 5(ln5)²/16≈0.81<1, average total 4 vs 6), so the tailored existence claim holds, though it is a dressed-up instance of the classical cobweb/lagged-feedback cycle in a decreasing one-dimensional map. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The arithmetic and stability check are right: equalization turns a constant-map economy into a locally attracting 2-cycle {1,3} with average total income 4 versus 6, an engineered illustration of the known principle that closing a lagged negative feedback loop can create flip oscillations. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The construction rigorously establishes the claimed attracting cycle and one-third long-run income loss, but it is a calibrated reduced-form example of classical decreasing-exponential dynamics, not a general result about redistribution. full review |
Two elastically coupled threshold faults can produce about 24% more earthquakes after tectonic loading is slowed on both faults, without changing their failure thresholds, stress drops, or elastic coupling.
supporting text
Consider an idealized stress-shadow model. Each fault fails when its normalized stress reaches 1 and then drops its own stress by 1. An A earthquake additionally reduces B's stress by 2; a B earthquake reduces A's stress by 0.1. Between events, tectonic loading increases stresses at rates bA and bB. Negative normalized stresses simply represent stress deficits below failure. These unequal normalized transfers need not violate elastic reciprocity: different failure stresses and slip increments can produce them. Their product is 0.2, compatible with a positive-definite reciprocal elastic stiffness matrix after restoring dimensional units. For bounded stress histories, long-run event rates satisfy rA + 0.1rB = bA and 2rA + rB = bB. Initially set bA = 1 and bB = 3. The resulting rates are rA = 0.875 and rB = 1.25, totaling 2.125. Now reduce both loading rates to bA = 0.5 and bB = 2.9. The rates become 0.2625 and 2.375, totaling 2.6375: an increase of approximately 24.1%. Both parameter sets satisfy bA > 0.1bB and bB > 2bA, ensuring that neither fault can be indefinitely suppressed in this inhibitory threshold construction. Slowing A removes enough large stress-shadow events to let B fail much more frequently. This is an existence result for a fixed-drop fault model, not a prediction that reduced tectonic loading generally increases seismicity; earthquake counts also should not be confused with seismic moment or released energy.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The arithmetic and the rate-balance argument are correct (total rate = −1.25 b_A + 1.125 b_B, giving +24.1%), but the effect is an elementary consequence of strongly asymmetric inhibitory coupling in an idealized model and the headline claim omits those restrictions. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: partly | The rate equations and the 24.1% figure check out (total rate = −1.25·b_A + 1.125·b_B, with bounded stresses verifiable), but the effect depends on highly non-proportional, independently set loading rates in an idealized inhibitory threshold model and is a standard disinhibition consequence of the inverse interaction matrix. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The 24.1% long-run event-count increase is mathematically valid for this reciprocal-elastic, fixed-drop threshold construction, but relies on unequal stress-loading reductions and strong asymmetric inhibition rather than a general earthquake response. full review |
A single-site receptor can have an arbitrarily narrow equilibrium temperature window for binding while its binding enthalpy and any prescribed finite number of temperature derivatives of that enthalpy vanish exactly at the window’s center.
supporting text
Consider an idealized receptor with many interconverting conformations but only one ligand-binding site. Choose an even integer N and a reference inverse temperature β₀=1/(RT₀). Its unbound conformations have enthalpies kE for even k between 0 and N; its bound conformations have enthalpies kE for odd k. Assign each conformation the temperature-independent entropy weight C(N,k)exp(β₀kE), with an additional common multiplier setting the affinity scale of the bound pool. Writing y=exp[−E(β−β₀)], the unbound and bound partition sums are proportional to [(1+y)^N+(1−y)^N]/2 and [(1+y)^N−(1−y)^N]/2. Consequently, K(T)/K(T₀)=(1−z^N)/(1+z^N), where z=tanh[E(β−β₀)/2]. Because N is even, T₀ is the affinity maximum. Near T₀, log K first changes at order (β−β₀)^N. The binding enthalpy, −∂log K/∂β, therefore vanishes together with its first N−2 temperature derivatives at T₀. Nevertheless, increasing E makes the affinity window arbitrarily narrow; outside it, K approaches zero. This is an equilibrium construction, not an experimentally established receptor. It requires finely matched conformational entropies. Its distinctive implication is that even several orders of locally null calorimetric information cannot exclude an extremely temperature-selective, noncooperative binding response.
| claude-fable-5.1 | correct with caveats | new as far as I know | fair: yes | The algebra and derivative-counting are correct and the construction does establish the existence claim, but only as a finely tuned equilibrium idealization with temperature-independent conformational enthalpies/entropies and unbounded energy scales. full review |
| claude-opus-5.5 | correct with caveats | new as far as I know | fair: yes | The binomial-parity construction rigorously gives K/K₀=(1−z^N)/(1+z^N) with ln K−ln K₀ ≈ −2(E(β−β₀)/2)^N and width ∝1/E, so the existence claim holds, though only for a finely tuned idealized ensemble with large hidden enthalpy spans. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The idealized binomial-parity construction rigorously gives arbitrarily narrow threshold-defined binding windows with any finite set of null enthalpy derivatives, but requires unbounded, finely matched thermodynamic parameters and does not establish experimental realizability or publication priority. full review |
Three one-parameter ridge learners trained on distinct fixed inputs can amplify an arbitrarily small prediction error without bound through cyclic exact distillation, even though every two-learner mutual-distillation subsystem eliminates all errors within two rounds.
supporting text
Let there be three inputs x₁, x₂, x₃ and three scalar models f_i(x) = a_i φ_i(x). Define the feature values by the following rows: at x₁ they are (1,2,0), at x₂ they are (0,1,2), and at x₃ they are (2,0,1). Indices below are cyclic. Learner i trains only at x_i, using learner i+1 as its teacher. Each synchronous round exactly minimizes, over its new coefficient a, the ridge objective (a−f_{i+1}(x_i))² + λa², with 0 < λ < 1. The unique update is therefore a_i(new) = [2/(1+λ)]a_{i+1}(old). Starting all coefficients at any positive ε makes every coefficient grow by the factor 2/(1+λ) each round. If the correct target function is zero, mean squared prediction error on the uniform distribution over the three inputs consequently diverges, despite beginning arbitrarily close to zero. Now retain any two learners and let them teach each other using exactly the same inputs, features, and ridge objectives. One teaching direction has multiplier 2/(1+λ), but the reverse direction has multiplier zero. The two-step update is therefore identically zero. Thus exhaustive testing of every pair misses a strictly three-way instability. The mechanism requires different training inputs: each learner extrapolates strongly precisely where its predecessor trains, closing an amplifying loop that no pair can close.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | The construction is verified: the cyclic update is a ← (2/(1+λ))·P·a with spectral radius >1 while every pairwise update matrix is nilpotent of index 2, so the claim holds exactly, though it is an elementary instance of cycle-gain instability rather than a new phenomenon. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: partly | The arithmetic checks out (cycle gain (2/(1+λ))³>1, each pair's update nilpotent), but it is a hand-built instance of the known fact that pairwise-stable subsystems need not make a stable cyclic system, and 'eliminates errors' relies on the zero target. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The construction is valid under synchronous exact ridge minimization with zero ground truth, and realizes a classical amplifying directed cycle whose every two-learner restriction is nilpotent. full review |
Two four-state Markov models with identical spectra and entropy rates can each contract every mean-zero signal by at least a factor of two per step, and their probability average can mix perfectly in one step, while their logit average mixes arbitrarily slowly.
supporting text
Choose 0<ε<1 and put c=(1+ε)/2. Define symmetric transition matrices by normalizing each row of the following weight matrices by 2+2ε: W_A = [[c,c,1,ε],[c,c,ε,1],[1,ε,c,c],[ε,1,c,c]], W_B = [[c,c,ε,1],[c,c,1,ε],[ε,1,c,c],[1,ε,c,c]]. Both chains are strictly positive and have uniform stationary distribution. They are related by a state permutation, so their spectra and stationary entropy rates agree. Their three nonconstant eigenvalues are 0 and ±(1−ε)/(2+2ε). Symmetry therefore makes their mean-zero Euclidean contraction factors less than 1/2. Their arithmetic probability average is exactly the matrix with every entry 1/4: all information about the initial state disappears after one transition. Instead average the logits log P_A(i,j) and log P_B(i,j), then normalize each row. Within either block {1,2} or {3,4}, the resulting unnormalized weights are c; between blocks they are √ε. The block-contrast vector (1,1,−1,−1) consequently has eigenvalue r = (1−√ε)²/(1+√ε)². As ε tends to zero, r tends to one and the mixing time diverges. Thus two equally fast, uniformly stationary constituent models—and an instantly mixing probability ensemble—can yield an almost disconnected logit ensemble. Geometric averaging preserves shared transitions but nearly erases complementary routes between the blocks.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | Every computation checks out (spectra 1, 0, ±(1−ε)/(2+2ε); arithmetic average = uniform matrix; geometric average has block eigenvalue ((1−√ε)/(1+√ε))²→1), giving a clean, correctly summarized illustration of the known fact that geometric/PoE averaging erases complementary transitions while arithmetic mixing unions them. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | All computations check out (spectra 0, ±(1−ε)/(2+2ε); arithmetic average = J/4; logit-average block eigenvalue (1−√ε)²/(1+√ε)² → 1), giving a valid explicit illustration of the known veto effect of geometric pooling. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The construction rigorously gives uniformly fast constituents and an exactly mixing arithmetic ensemble while the normalized geometric ensemble has mixing time diverging like epsilon to the minus one-half, though the exact example's publication priority is unverified. full review |
Centered, real-analytic, uniformly strongly log-concave densities can have exactly the same integral of every positive power as the standard Gaussian and variance below 1.454, yet have arbitrarily large Fisher information.
supporting text
Uniform strong log-concavity here holds separately for each density; no common curvature bound is asserted. Let φ be the standard Gaussian density. For 0<a<1 and δ>0, set c(t)=a(√(t²+δ²)−δ) and define the increasing analytic coordinate change x=t+c(t). Define p(x)=φ(t), without a Jacobian factor. Because c' is odd, substitution shows that ∫p(x)^q dx=∫φ(t)^q dt for every q>0, including normalization at q=1. Translation to mean zero preserves all these identities. For V=−log p, direct differentiation gives V''(x)=[1+a t³/(t²+δ²)^(3/2)]/[1+a t/√(t²+δ²)]³. This is positive and bounded above and away from zero for each fixed a,δ. Thus the density is real-analytic and uniformly strongly log-concave. Its Fisher information is I(p)=∫φ(t)t²/[1+c'(t)]dt. Let a approach one and choose δ=1−a. On any fixed negative interval bounded away from zero, the denominator tends uniformly to zero. Therefore I(p) tends to infinity. The variance nevertheless stays small. Layer-cake integration over the superlevel intervals [c(r)−r,c(r)+r] gives Var(X)=1+Var(c(R)), where R has density 2r²φ(r) on r>0. Since c is a-Lipschitz, Var(c(R))≤a²Var(R)=a²(3−8/π). Consequently 1≤Var(X)<4−8/π<1.454. Thus even the complete collection of power integrals, with tightly controlled variance, leaves Fisher information unbounded within this smooth convex class.
| claude-fable-5.1 | correct | special case or variant of known results | fair: partly | The construction is verified in full (power integrals, analyticity, per-density strong log-concavity, Var(X)=1+Var(c(R))<4−8/π, and I(p)≈1/(2δ)→∞), being an explicit equimeasurable-rearrangement example whose only ambiguity is that 'uniformly' does not mean a common curvature constant. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The level-set-shifting construction is verified in every step (power integrals, curvature bounds, Fisher information blow-up, variance below 4−8/π), with curvature bounds holding per density but not uniformly across the family, as the text states. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The construction rigorously proves the claim, even with a common curvature lower bound of 1/4, and is an analytic smoothing of the classical split-normal example. full review |
An ideal equilibrium model of a single ampholyte that aggregates only as far as tetramers can cross neutral pH exactly three times during dilution with pure water, without added salts or changes in temperature.
supporting text
Here is an explicit mass-action construction, not a claim about an already characterized compound. Let M be the neutral monomer, c* a concentration scale, u=[M]/c*, and y=[H+]/sqrt(Kw). Allow singly charged aggregates M_nH+ and M_n−H− for n=1,2,3,4, with equilibrium concentrations [M_nH+]=c* a_n u^n y and [M_n−H−]=c* b_n u^n/y. These expressions correspond to ordinary aggregation and protonation equilibria. Choose a=(1,12,1,2) and b=(7,1,7,1). All equilibrium constants are positive. At neutrality, water contributes zero net charge, while the ampholyte contributes c*[u^4−6u^3+11u^2−6u]=c*u(u−1)(u−2)(u−3). Thus neutral solutions occur at precisely three positive free-monomer concentrations. Their analytical ampholyte concentrations, counting monomer units, are C=c*[u+8u+26u^2+24u^3+12u^4], giving C/c*=71, 506, and 1881. These are genuine crossings, not merely three neutral algebraic states. Ideal-mixture free energy is strictly convex under the material and charge constraints, so equilibrium is unique. Moreover, proton-charge susceptibility at fixed ampholyte inventory is positive: eliminating the monomer chemical potential gives a positive Schur complement of the species susceptibility matrix. Because all three polynomial roots are simple and C increases strictly with u at neutrality, the equilibrium pH changes sides at each threshold. The engineered ingredient is alternating charge preference among successive aggregation orders.
| claude-fable-5.1 | correct | new as far as I know | fair: yes | The mass-action construction checks out (quartic roots u=1,2,3, C/c*=71,506,1881, and charge-susceptibility positivity verified via Cauchy–Schwarz), so an ideal tetramer-limited ampholyte model can indeed cross pH 7 exactly three times on dilution, though the example is contrived rather than chemically motivated. full review |
| claude-opus-5.5 | correct | new as far as I know | fair: yes | The algebra checks out (P(u)=u(u−1)(u−2)(u−3), C/c*=71, 506, 1881), and uniqueness of equilibrium makes C strictly increasing in u, so a tuned ideal tetramer-limited ampholyte model crosses neutral pH exactly three times on dilution, the maximum possible in this setting. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The constructed ideal mass-action model rigorously has exactly three neutral-pH crossings over a sufficiently broad dilution range, although it establishes neither a real compound nor publication priority. full review |
My own check. I re-did the algebra (net charge at neutrality u(u−1)(u−2)(u−3); C/c* = 71, 506, 1881 at u = 1, 2, 3) and simulated the full equilibrium, including water's own charge, while diluting from C/c* = 5000 to 5 (c* = 1 mM). The pH crosses 7 exactly three times, at C/c* ≈ 1881, 506 and 71, as claimed. Caveat the text omits: the pH stays within about 0.003 units of 7 at this scale, so the crossings are real but practically invisible.
For any prescribed integer n, a linear slab-ocean energy-balance model can be constructed with strictly positive seasonal radiative damping and a fixed, seasonally sign-changing forcing of positive annual mean, such that its annual-mean temperature response reverses sign at least n times as mixed-layer heat capacity increases.
supporting text
Use nondimensional time with annual period 2π and consider C dT/dt = F(t) − λ(t)T. Here C is mixed-layer heat capacity, λ(t) = 1 + εΣ cos(jt), and F(t) = εb + Σ fⱼ cos(jt), with both sums running from j = 1 to n. Choose b > 0 and ε sufficiently small that λ remains positive throughout the year. Every model therefore has a unique stable periodic response. A perturbation expansion of that response gives ⟨T⟩/ε = b − (1/2)Σ fⱼ/(1+j²C²) + O(ε). The correction comes from covariance between seasonal damping and the temperature oscillations; changing C alters the attenuation of each harmonic differently. Choose distinct positive numbers r₁,…,rₙ. The rational function R(x) = b∏(x−rₖ)/∏(x+1/j²) has exactly those positive simple zeros and tends to b at infinity. Partial fractions express it as b + Σ dⱼ/(1+j²x). Setting fⱼ = −2dⱼ therefore makes the leading annual response equal R(C²). For sufficiently small fixed ε, its alternating signs on either side of every root persist in the exact solution. The forcing and damping schedules remain fixed while only C changes. This is a deliberately constructed seasonal-forcing result, not a claim about everywhere-positive greenhouse forcing.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | The perturbation expansion, partial-fraction construction, and persistence argument all check out, so the claim holds as a deliberately engineered corollary of standard linear seasonal-EBM rectification theory. full review |
| claude-opus-5.5 | correct | new as far as I know | fair: yes | The perturbation formula, partial-fraction construction and small-ε persistence argument all check out, giving at least n sign reversals of annual-mean temperature with C for a fixed, deliberately tuned seasonal forcing and damping. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The arbitrary-number-of-reversals existence claim is rigorously valid for deliberately constructed, fixed sign-changing forcing, although publication novelty of the exact climate formulation is unverified. full review |
Under credit awarded to the publication that first completes a discovery, two papers that are equally insufficient alone and equally dispensable in the finished archive can differ quadratically in expected credit merely because they belong to alternative validation routes of different lengths.
supporting text
Consider a discovery with two sufficient validation routes. One requires papers a and c together. The other requires paper b together with m additional papers d₁, …, dₘ. Neither route has any sufficient proper subset. Assume all m + 3 papers eventually appear, in uniformly random order, and priority credit goes to whichever publication first completes either route. Paper a and paper b agree on two familiar retrospective assessments. Neither establishes the discovery alone. Removing either from the completed archive leaves a complete alternative validation route. Both therefore have zero standalone success and zero indispensability to the final archive. Their expected priority credit nevertheless differs sharply. Represent publication order by independent uniform arrival times. Conditional on a arriving at time t, it wins precisely when c has already arrived but the entire competing route has not. Its winning probability is the integral of t(1 − t^(m+1)), namely (m + 1)/(2(m + 3)). For b, the corresponding integral is t^m(1 − t²), giving 2/((m + 1)(m + 3)). Their credit ratio is exactly (m + 1)²/4. This does not establish that priority credit is unjust. It establishes that large credit differences can arise without differences in standalone sufficiency, final indispensability, or publication-speed distributions. The architecture of alternative validation routes is itself a powerful generator of historical prominence.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The computation is correct (it is the Shapley–Shubik pivotal probability for a simple game with two disjoint minimal winning coalitions of sizes 2 and m+1, giving ratio (m+1)²/4), but 'quadratically' holds for the ratio not the difference, and the result depends on uniform ordering, disjoint routes, and winner-take-all credit. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The integrals and the (m+1)²/4 ratio check out exactly. The result is an instance of Shapley–Shubik power indices differing among players with equal standalone and grand-coalition marginal value, under the stated uniform-random-order model. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The quadratic credit ratio is rigorously correct for exactly two disjoint routes under uniformly random publication order, and is a direct application of Shapley–Shubik pivotality. full review |
For a simple acyclic parity model, implementing all surgical interventions by weighted equation-violation penalties requires exponentially separated energy scales if every intervention must retain a fixed positive energy margin.
supporting text
Consider n binary variables with equations x_i = XOR(x_1,...,x_{i-1}), taking the empty XOR as zero. Suppose an implementation uses the energy E(x) = Σ_i w_i 1[x_i ≠ XOR(x_1,...,x_{i-1})], with positive weights. Interventions clamp selected variables without removing their energy terms. Require the correct surgical solution to beat every competitor by at least Δ > 0, uniformly over interventions. Fix i. Clamp all earlier variables to zero and leave only x_i free. Clamp each later variable, successively, to the opposite of its prescribed value calculated with x_i = 0. With x_i = 0, equation i is satisfied and every later equation is violated. Flipping x_i to one violates equation i but satisfies every later equation: each later parity includes x_i, while its other inputs remain clamped. The required margin therefore forces w_i ≥ Σ_(j>i) w_j + Δ. For the final variable it forces w_n ≥ Δ. Backward induction gives w_i ≥ 2^(n-i)Δ, and hence w_1 ≥ 2^(n-1)Δ. These bounds are attainable using precisely those weights. Thus an n-step causal computation with elementary Boolean equations can require an exponential penalty hierarchy in this particular equilibrium implementation. Matching the unmanipulated solution is cheap; uniformly matching its intervention semantics imposes the large resource requirement.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The derivation is correct and tight (wᵢ ≥ 2^{n−i}Δ, attainable), but the exponential hierarchy is an artifact of retaining clamped variables' penalty terms in a dense all-predecessor parity model, and the argument is essentially the classical lexicographic-weight bound. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: partly | The derivation w_i ≥ Σ_{j>i} w_j + Δ (hence w_1 ≥ 2^{n-1}Δ, and w_1/w_n > 2^{n-2}) is correct and geometric weights provably suffice, but the result hinges on retaining penalty terms of clamped variables and on a complete-DAG, unbounded-fan-in parity model, the first of which the one-line claim omits. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The exponential bound is exact for fixed retained equation penalties and unrestricted hard interventions, and its mechanism is standard superincreasing-weight encoding of lexicographic priorities. full review |
In a parameter-free family of multiplicative hidden-factor theories with identical observable means and variances, increasingly precise measurements of an exactly null signal drive posterior belief beyond every finite bound on the number of hidden factors, under any prior giving every finite count positive probability.
supporting text
Let theory T_n posit n independent hidden factors U_i, each uniform on [−1,1], and predict the single observable signal Y=3^(n/2)∏U_i. Every theory has mean zero and variance one. There are no adjustable parameters. Measurements successively establish that this same signal lies within ε of zero; they are not independent redraws. For sufficiently small ε, put t=ε/3^(n/2). Direct integration of the product distribution gives P(|Y|≤ε | T_n)=t Σ[k=0 to n−1] (log(1/t))^k/k!. For each fixed n, its leading dependence is therefore a positive constant times ε(log(1/ε))^(n−1). Consequently, the likelihood ratio of T_(n+1) to T_n diverges as measurement precision increases. More strongly, fix any finite bound K. The combined likelihood-weighted prior mass of T_1 through T_K grows more slowly than the mass contributed by T_(K+1) alone. Hence the posterior probability that n≤K tends to zero, whatever proper prior was used, provided every n initially had positive probability. Additional factors produce a sharper concentration near zero, compensated by rare large signals that preserve the variance. Thus better null measurements can generate indefinitely escalating ontological commitment without identifying any particular factor. This is a misspecification warning, not a claim that infinitely many factors have been discovered: an exactly zero signal has probability zero under every candidate, and the conclusion concerns the limit of ordinary finite-precision updates, not conditioning directly on that null event.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The computation is right (Gamma(n,1) tail gives P_n ~ ε(log 1/ε)^{n−1}, so posterior mass escapes every finite K as ε→0), but the claim silently requires a proper prior and relies on a family whose densities become arbitrarily singular at zero, a known type of Bayesian model-comparison pathology rather than a general theorem. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The formula, the divergence of likelihood ratios, and the posterior escape past every finite K are all rigorously correct under nested interval data (and survive Gaussian noise), though the effect is only logarithmic in precision and hinges on the density singularity at an exactly null signal. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The posterior-escape theorem is rigorously correct for the specified proper-prior, single-signal interval experiment, but is a classical boundary-concentration phenomenon rather than a general result about hidden-factor theories. full review |
For any fixed r ≥ 2, two experiments can be equally informative for every decision problem involving at most r rival theories, while one identifies the true theory with probability approaching one and the other with probability approaching zero when the entire finite theory set is considered.
supporting text
Here “equally informative” means Blackwell-equivalent: either experiment can simulate the other without knowing which theory is true. Use a uniform prior over N theories. An experiment draws an integer K independently of the truth, then reports a uniformly selected K-element subset containing the true theory. Its optimal identification probability is E[1/K]. Take N = M^r and nodes x_j = M^j, for j = 0,…,r. Define signed weights w_j = 1/∏_{l≠j}(x_j−x_l). These weights annihilate every polynomial of degree at most r−1. Normalizing their positive and negative parts therefore produces two probability distributions on K with matching moments through degree r−1. As M grows, the absolute weights at x_0 = 1 and x_1 = M become asymptotically equal and dominate all remaining weights. Thus one distribution concentrates at K = 1, while the other concentrates at K = M. Their identification probabilities consequently approach one and zero. Why does every small comparison miss this difference? Restrict attention to a menu of s ≤ r theories. Conditional on the truth belonging to that menu, probabilities for the reported subset’s intersection with the menu are expectations of polynomials in K of degree at most s−1. They therefore agree exactly between experiments. Given that intersection, the rest of the report contains no further information about which remaining theory is true and can be simulated separately. Thus even complete decision-theoretic agreement on every bounded-size comparison need not determine whether an experiment resolves the overall underdetermination.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The construction and argument check out rigorously (menu statistics are degree-(s−1) polynomials in K, divided-difference weights give moment-matched distributions concentrating at K=1 vs K=M), but the claim hides that N→∞, that menus are compared conditionally on the truth lying in them, and that the prior is uniform. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The moment-matching construction is valid: restriction to any ≤r-theory menu yields identical sufficient-statistic laws (degree ≤ r−1 polynomials in K), while E[1/K] tends to 1 versus 0 as N = M^r grows. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The construction rigorously gives exact Blackwell equivalence on every fixed-size theory menu alongside opposite global identification limits, provided the theory set grows and equivalence is understood menu by menu. full review |
In a persistent lab-or-method artifact model, a replication introducing neither a new laboratory nor a new method can provide the same leading-order evidential gain as introducing both, while arbitrarily many new laboratories can provide no such gain.
supporting text
Represent a replication program by a bipartite graph: laboratories are vertices on one side, methods on the other, and completed successful experiments are edges. Under H, every experiment succeeds. Under the artifact rival, each laboratory and each method independently carries a persistent artifact with probability ε; an experiment succeeds exactly when its laboratory or its method carries an artifact. All experiments succeed precisely when the artifact-bearing vertices cover every edge. As ε approaches zero, the rival's likelihood is therefore cε^τ + O(ε^(τ+1)), where τ is the minimum vertex-cover size and c counts minimum covers. The Bayes factor consequently has order ε^−τ. For bipartite graphs, τ equals the maximum matching size. Now start with two disconnected two-leaf stars: laboratory A uses methods x and y, while laboratories B and C use method z. The maximum matching has size two. Add a successful experiment by B using x. A matching of size three now exists: A–y, B–x, C–z. This entirely cross-reused replication raises the evidential exponent by one, exactly as would an isolated experiment introducing both a new laboratory and a new method. By contrast, adding any number of new laboratories using z leaves the exponent at two. Within this explicit model, evidential novelty is neither headcount nor a count of distinct methods: it is an augmenting structure in the replication network.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: yes | Within the idealized artifact-cover model the derivation is correct (König's theorem gives τ = max matching, and the worked example checks out), with the caveats that equality holds only in the leading exponent, not the constant, and the 'no gain' case depends on the chosen attachment to an already-covered method. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: partly | The example checks out (τ goes 2→3 via the cross-reused edge B–x, and stays 2 under any number of new labs on z), but 'same leading-order gain' holds only for the ε-exponent/log-evidence (constants c=4 vs 2), and it relies on ε→0 and equal artifact rates. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: partly | The construction rigorously establishes equal increases in evidential exponent, not equal leading Bayes-factor coefficients, using classical vertex-cover and matching results. full review |
For an unknown strictly increasing analytic instrument response arbitrarily close to the identity, infinitely many exact measurements under known shifts in an arbitrarily short interval identify the signal, although at a perfectly linear response no finite collection identifies it unless a reading hits the calibrated zero.
supporting text
Let the unknown signal be θ. Applying a known shift t produces the exact reading f(θ+t). The response f is real analytic on the real line, strictly increasing, satisfies f(0)=0, and obeys |f(x)−x|≤ε everywhere, for a fixed ε>0. These are restrictions on the instrument, not a known calibration curve. First suppose the actual response is f(x)=x. Take any finite collection of shifts whose readings are nonzero. Replace θ by θ+δ. For sufficiently small nonzero δ, one can construct an analytic perturbation h satisfying h(0)=0 and h(θ+δ+t_i)=−δ at every sampled shift. Finite interpolation with Gaussian radial basis functions supplies such an h, with both its supremum norm and its derivative supremum norm tending to zero with δ. Thus g(x)=x+h(x) remains strictly increasing and ε-close to the identity, yet reproduces every reading under the different signal. A zero reading is exceptional: strict monotonicity and f(0)=0 immediately determine θ. Now take infinitely many distinct shifts accumulating inside any short interval. If two candidate pairs reproduce those readings, analyticity forces f(θ+t)=g(θ′+t) for every real t. Substituting t=−θ′ gives f(θ−θ′)=0, hence θ=θ′. The distinction is exact identifiability, not practical stability: analytic continuation need not tolerate measurement noise.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | Both halves hold—identity theorem plus f(0)=0 and monotonicity give exact identifiability from accumulating shifts, while Gaussian RBF perturbations show finite nonzero readings never pin down θ at the identity response—but it is a routine application of classical analytic-continuation facts rather than a new result. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | Both halves check out: the identity theorem gives exact identification from shifts accumulating in a bounded set, and small Gaussian-RBF perturbations show that at the identity response no finite set of nonzero readings identifies θ. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: partly | The theorem is valid under the stated anchored, globally real-analytic model with positive instrument uncertainty, but its headline needs those qualifications and its proof combines classical identity and interpolation results. full review |
Any upward-closed pattern of which instrument combinations identify a scalar quantity can be realized by a finite noiseless linear measurement system, even when every reading has nonzero direct sensitivity to that quantity.
supporting text
Let instruments be indexed by a finite set, and prescribe any upward-closed family of nonempty subsets as the combinations that should identify θ. Call its inclusion-minimal members sufficient groups. For each sufficient group S containing k instruments, introduce a separate block of readings and k−1 unrestricted nuisance offsets z₁,…,zₖ₋₁. Give its first k−1 instruments readings y_j = θ/k + z_j. Give the last instrument the reading y_k = θ/k − Σz_j. Every reading has direct sensitivity 1/k to θ. An instrument belonging to several groups returns several channels, with separate nuisance offsets for each block. Whenever all members of S are available, their block sums exactly to θ. Hence every prescribed sufficient combination identifies the target. Conversely, consider a combination containing no complete sufficient group. In each block, at least one reading is missing. For any proposed value of θ, the observed partial block can be reproduced by adjusting its nuisance offsets: if the final reading is missing, choose the observed offsets directly; otherwise an offset belonging to a missing earlier reading absorbs the remaining discrepancy. Blocks have separate offsets, so these adjustments are mutually compatible. Every real θ remains possible. Thus linearity and genuine target sensitivity impose no diminishing-returns law on measurement combinations. Redundancy, complementarity, and abrupt gains from adding one final method can follow any monotone pattern; they cannot generally be inferred from instruments considered individually or in pairs.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The construction is valid (verified via a ∉ col(B) block-diagonal argument) and realizes any upward-closed family of nonempty subsets, but it relies on vector-valued instruments and unrestricted real nuisances, omits trivial edge cases, and is essentially additive secret sharing over minimal qualified sets recast as a linear identifiability statement. full review |
| claude-opus-5.5 | correct with caveats | known | fair: partly | The construction is valid and realizes any monotone identification pattern (it is the Ito–Saito–Nishizeki/Benaloh–Leichter additive secret-sharing scheme over the reals), but it needs multi-channel instruments and unrestricted nuisance offsets, and its 'nonzero direct sensitivity' is a parametrization artifact rather than genuine sensitivity. full review |
| gpt-6.1-sol | correct with caveats | special case or variant of known results | fair: yes | The construction rigorously realizes every finite monotone identification pattern with multichannel instruments and unrestricted nuisances, but it is a variant of standard additive secret-sharing constructions rather than a new general theorem. full review |
A theory with just one continuously adjustable auxiliary parameter can fit every proper subset of an arbitrarily large dataset exactly while remaining separated by a positive error margin from the complete dataset.
supporting text
Choose any integer n≥2. There are n measurement settings, and the observed result at every setting is zero. Theory A predicts these zeros without adjustable parameters. Theory B has one shared auxiliary θ in the compact interval [1,n] and predicts, at setting i, f_i(θ)=∏_{j≠i}(θ−j)². Delete observation k. Setting θ=k makes every remaining prediction zero, because each remaining polynomial contains the factor (θ−k)². Thus every proper subset of the dataset fits B exactly: choose k outside that subset. Nevertheless, no θ fits the complete dataset. At a noninteger θ every prediction is positive; at an integer k, prediction k is positive. This discrepancy is robust, not merely an objection about exact equality. The continuous function g(θ)=max_i f_i(θ) is positive throughout a compact interval, so it has a strictly positive minimum. Measurement tolerances below that minimum exclude B using the complete dataset, while every proper subset still fits exactly. Consequently, auxiliary-parameter count alone places no upper bound on the number of measurements needed for a minimally decisive refutation. Here the unbounded resource is the algebraic complexity of the parameter's effects, not the number of adjustable parameters. Arguments that infer evidential redundancy merely from having many more measurements than auxiliaries therefore need additional structural assumptions.
| claude-fable-5.1 | correct with caveats | special case or variant of known results | fair: partly | The construction is valid (for each n a one-parameter polynomial family fits every proper subset exactly yet misses the full all-zero dataset by a uniform positive margin, even without the compactness restriction), but it is a tailored, n-dependent, high-degree family illustrating the known point that parameter count is uninformative without constraints on functional form. full review |
| claude-opus-5.5 | correct with caveats | special case or variant of known results | fair: yes | The construction is valid (every proper subset fits at θ=k, the full set has a strictly positive minimax misfit), but the theory must be rebuilt with growing polynomial degree for each n, and the result is a Helly-number/VC-style illustration rather than a new finding. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The construction rigorously proves the finite existence claim, but its mechanism is a classical nonconvex intersection example realized by increasingly complex polynomial predictions. full review |
Two finite theory–record pairs can agree on the number and size of minimal falsifiers, each reading's participation in them, the maximum number of disjoint falsifiers, and the minimum repair count, yet have exponentially different probabilities of surviving the same random withholding audit.
supporting text
Take n binary readings, where n is divisible by twelve, and suppose every recorded value is one. Define a theory using a graph on the readings: adjacent readings may not both equal one. A minimal falsifier is therefore an edge whose two endpoints were observed as one. For theory A, use disjoint four-cycles; for theory B, use disjoint six-cycles. Both have exactly n minimal falsifiers, all of size two. Every reading participates in exactly two. Both have n/2 mutually disjoint minimal falsifiers. Both require changing or deleting at least n/2 readings to repair the complete record, because that is their minimum vertex-cover size. Now independently retain each reading with probability one-half. The audit finds no refutation precisely when the retained vertices form an independent set. A four-cycle has seven independent sets among its sixteen vertex subsets. A six-cycle has eighteen among sixty-four. Consequently, the survival probabilities are (7/16)^(n/4) and (18/64)^(n/6). Their ratio is (343/324)^(n/12), which grows exponentially. Thus even this collection of local severity measures and global repair measures does not determine resilience under random withholding. The missing information is how minimal falsifiers overlap beyond their individual participation counts. This construction concerns survival of an audit, not posterior truth: its methodological implication is that auditing resilience cannot generally be inferred from counts of anomalies, independent contradictions, and required repairs alone.
| claude-fable-5.1 | correct | special case or variant of known results | fair: partly | The C4-vs-C6 construction is verified in every step (equal counts, degrees, matching and vertex-cover numbers; independent-set counts 7 and 18 giving survival ratio (343/324)^(n/12)), so the claim holds, though it repackages the known fact that independence polynomials are not fixed by such invariants and 'exponentially different' refers to the ratio of two vanishing probabilities. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | The construction checks out (and the gap holds for every retention probability p), giving exponentially different audit-survival probabilities despite equal falsifier counts, degrees, matching and vertex-cover numbers; mathematically it is a known independence-polynomial/Kahn–Zhao-type phenomenon, reframed for falsification. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | The construction rigorously proves exponential separation in audit-survival ratios despite agreement on all listed statistics, using standard independent-set counting. full review |
For additive, state-independent measurement error, exact mean calibration on two arbitrarily nearly linear scales can force the instrument to be noiseless, although calibration within any positive tolerance on those same scales permits arbitrarily large error variance.
supporting text
Let the reading be Y = θ + E, where E has a fixed distribution independent of θ and finite first moment. Choose two reporting scales f₁(x) = x + a sin x and f₂(x) = x + a sin(√2 x). Taking a sufficiently small makes both scales strictly increasing and arbitrarily close to the ordinary scale, uniformly in both value and derivative. Suppose E[fⱼ(θ + E)] = fⱼ(θ) for every θ and both j. Comparing the constant, sine, and cosine coefficients forces E[E] = 0, E[cos E] = 1, and E[cos(√2 E)] = 1. Since cosine never exceeds one, these last equalities require E to belong almost surely to both 2πZ and √2πZ. Their intersection is {0}. Thus exact calibration entails zero noise, without relying on convexity of either scale. Yet this conclusion has no tolerance-stable counterpart. Take E equally likely to be ±2πn. Calibration on f₁ is exact, while the maximum calibration error on f₂ is a|cos(2π√2 n) − 1|. There are arbitrarily large integers n making this smaller than any prescribed positive tolerance, while Var(E) = 4π²n² diverges. Consequently, two innocuous-looking calibration conventions can jointly impose perfect measurement as an exact mathematical constraint, while every finite-tolerance implementation of those conventions remains compatible with arbitrarily noisy instruments.
| claude-fable-5.1 | correct | special case or variant of known results | fair: yes | The argument is valid: exact calibration on x + a sin x and x + a sin(√2x) forces φ_E(1)=φ_E(√2)=1 hence E=0 a.s., while ±2πn errors with √2n nearly integral give arbitrarily good tolerance calibration with divergent variance; it is a clean application of standard characteristic-function and Diophantine facts rather than a new result. full review |
| claude-opus-5.5 | correct | special case or variant of known results | fair: yes | Both the exact-degeneracy argument (cosine/characteristic-function lattice intersection at incommensurable frequencies) and the Dirichlet-approximation counterexample with unbounded variance check out; the content is a measurement-theoretic repackaging of classical facts. full review |
| gpt-6.1-sol | correct | special case or variant of known results | fair: yes | For fixed nonzero sinusoidal perturbations with incommensurate frequencies, exact transformed-mean calibration forces zero noise, while Pell approximations yield uniformly vanishing calibration errors with diverging finite variances. full review |