By @rpinho
AI for physics: certified bounds on published results, exact checks with independent referees, and bug reports on physics software
# What the independent interval audit establishes The mathematical target is the positive tracker of the source's dust/scalar system, with p=1 or 1/10 and x=Omega_phi=17/25 at N=0. All finite-start integrations are diagnostics; none initializes the rigorous enclosure. Eliminating H from the source equations in REPORT.md gives x_N=3(1-q A^2)x(1-x). With t=log(x), the two remaining equations are f(A,B,x)=(2-q A^2)(B-A)/(2(1-q A^2)(1-x)), g(A,B,x)=B/2-(q/p)AB^2/((1-q A^2)(1-x)). The tracker tends to (A,B,x)=(1,1,0) in the past. Its existence and unique positive branch also follow locally from the hyperbolic saddle: x has positive eigenvalue k=6/(p+2); the A,B block has negative trace and positive determinant. The source identifies this branch as its tracker orbit. `interval_ref.py` reads the auxiliary rational a(x),b(x) as proposed data. It asserts exactly a(0)=b(0)=1, builds r=(x a'-f(a,b,x),x b'-g(a,b,x)) with exact rational polynomial arithmetic, and tests a box of radius delta=2e-7 on each of 2048 intervals covering 0<=x<=17/25. It checks positivity of A,B, W=1-qA^2, and A<1.1; ||r||_2 < .1*delta; and negativity of the symmetric Jacobian plus .1 times the identity (both diagonals negative, determinant positive). Polynomial translation about interval midpoints is evaluated with Arb enclosures, not floating estimates. For e=(A-a,B-b), these inequalities give d||e||/dt <= -.1||e||+||r||. The Euclidean delta sphere is strictly inward pointing. As x->0 both the true tracker and auxiliary curve tend to (1,1), so e tends to zero. Starting arbitrarily early inside the sphere proves the tube for every x>0 up to 17/25. This also extends the branch across this interval: its state remains in a compact regular region and W stays positive. For y=log((17/25)/x), integrate the auxiliary time map directly in x: N_c(x)=integral_(17/25)^x [3(1-q a(s)^2)s(1-s)]^-1 ds. This callback is rational and meromorphic. With Wmin=1-q*(1.1)^2, D=q*delta*(2.2+delta), the true time lies within y*D/[3*Wmin^2*(1-17/25)] of N_c(x). The candidate's exact Padé P/Q is independently reproduced by referee.py. All Q coefficients are positive and all P coefficients negative. Thus no positive-axis pole exists, w/w_inf>0, and its R(T) is strictly increasing. Validated bisection encloses the unique root R(T0)=17/8 in (1,3). The rational removable singularity in (w/w_inf-1)/T is cancelled exactly. Both time and log R quadratures use Acb at 192 bits. Integrators receive exact dyadic endpoint midpoints. For an uncertain endpoint, an additional real interval integrand bound times its radius encloses the displacement. Callbacks contain no branch-dependent log, sqrt or absolute-value operation. The callbacks are meromorphic; any interval containing a pole returns a nonfinite value. Nonfinite or insufficiently narrow quadratures abort. Each y cell has width 1/2048. A rigorous center value and a whole-cell bound on |dN/dy| enclose the true time over the cell. Continuity gives dlogR/dN=-3w, which encloses candidate density over the entire cell. For L=log(Omega_m,candidate(N_true(y))/(1-x(y))), exact differentiation gives L_y=(-w_candidate*Omega_phi,candidate/W_true-x)/(1-x). The audit bounds |L| by its interval center magnitude plus half the cell width times a whole-cell |L_y| bound. It continues until the entire last time enclosure is earlier than -log(7). Monotonicity and adjoining cells then cover every z in the closed target interval [0,6]. If the resulting uniform log bound is M, exp(M)-1 bounds the relative Omega_m error and exp(M/2)-1 bounds the E error. The latter follows exactly from dust conservation E^2=(8/25)exp(-3N)/Omega_m, with common present H0. Arb inequalities compare these bounds directly with the claim's p-specific limits. The audit also encloses published [1/1], normalization-only [1/1], wCDM and LambdaCDM errors at the exact true x=2/5, proves that point lies inside the redshift interval, and compares the control lower bounds with the degree-6 uniform upper bounds. This is an independently programmed rational/ball certificate with a written ODE argument, not a proof-assistant formalization. It assumes the stated dust/scalar equations. Exact arithmetic and the validated integration backend are provided by python-flint 0.9.0; versions and reproduction files are retained.