Independent mechanical audit, 2026-11-02. Final verdict: see VERDICT.md. The scoped printed-model theorem survives the audit. I found no false step in the global convergence, fold, or hyperbolic basin classification. This is a mathematical verification for exactly frozen decimal rates, not an empirical performance comparison. I read the candidate proof and its certificate summaries, but did not read, import, or execute its search, construction, or verifier code. The referee programs here were written independently from the primary equations and the mathematical assertions being tested. Primary-source consistency The live [Europe PMC XML](https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7079339/fullTextXML) is byte-identical to the frozen primary input: 352197 bytes, SHA256 `44a73bed7dbeaef770b73c291dc2fe5319b99ed578be4796e3afa165685500f8`. Its byline is Victor Garcia, Sebastian Bonhoeffer and Feng Fu; [PubMed](https://pubmed.ncbi.nlm.nih.gov/32035826/) independently agrees. I inspected printed Eqs.21–23, the complete NK/CTL section, Appendix C, and Table 3. The table's saturation constants give precisely the claimed rational n and e, with a=257/500, b=51/50000000000 and omega=11/100000000. Scaling y=N/sigma and z=E/(sigma*omega) gives the claimed U=c*sigma and V=k*sigma*omega with no time rescaling. The [official Figshare catalog](https://api.figshare.com/v2/articles/11536824) and directly downloaded `analysis_FengNKCTL.R` also match their frozen versions; the latter has MD5 `5b50e1ed1c7c4b2022b3c5ce366fc408`. Its NK nullcline uses `hg + T_w^2`, while printed Eq.22 uses kappa_bn+T. It references an absent parameter helper. Consequently the unmodified public R script is not an executable reproduction of this printed model. The candidate explicitly excludes that code variant, correctly. At T=10000 the printed n is about 1.15909, whereas that R formula gives about 1.90894; the difference is material. What the paper supplies, and what is added Appendix C eliminates N and E to obtain a rational nullcline with a quintic positive-equilibrium numerator. It explicitly distinguishes the nullcline's slope from full three-dimensional stability and uses numerical bistability examples. Its statement that the general model may have five nonzero fixed points is not an attained five-root example at frozen Table 3 rates. It does not give the audited global convergence theorem. Relative to that paper, the candidate adds a fixed-rate absorbing-tail energy certificate, global convergence including degenerate parameters, the exact physical fold classification, and the interior basin-boundary theorem. The earlier parent count result is a mathematical input, which I rechecked by a different method; this report does not reopen its separate taste decision. Independent exact reconstruction `independent_audit.py` uses SymPy rational polynomial arithmetic. It reconstructs every rate from the printed summands and differentiates the original three-state vector field. It verifies the p and r equations, the complete energy derivative, both endpoint damping inequalities over the entire half-line, and the exact eigenvalue margin 125/7757. The two cleared numerator polynomials have degree 13 and strictly positive coefficients. Coefficients, not just PASS labels, are retained in `pod_evidence/independent_exact.json`. For the equilibrium-count input I independently rebuilt the quintic and enumerated all 16 strict coefficient-sign patterns with more than two variations. Exact strict Fourier–Motzkin elimination in U,V proves each infeasible for U,V>0. Points with zero coefficients are covered by arbitrarily close generic parameters preserving every nonzero sign: inserting signs cannot reduce sign variation. Descartes therefore gives at most two positive roots counted with multiplicity, without relying on the claimant's threshold-ordering calculation. Solving F=F_T=0 independently reproduces both rational fold formulas. An exact Descartes/sign argument isolates its endpoint, and a positive-coefficient interval transformation proves Vsn'<0 on [0,2200]. The tangent identity then proves Usn'>0. The endpoint is Tcrit approximately 2132.6022002907025, Umax approximately 0.7283954288844874; Vcrit is exactly 10922486893/12500000000000. I checked the full equilibrium Jacobian cubic, its determinant identity, both positive Routh margins, the logarithmic energy coefficient at zero, and the removable derivative at U=a. These are exact checks; decimal endpoint values are descriptive only. Analytic steps audited separately from the programs 1. Positivity and finite-time existence: logistic comparison bounds T. The positive lower rate bounds on n and e then bound y and z. T=0 is invariant and converges directly to (0,1,0). Positive T cannot reach zero in finite time. 2. Absorption: n<=3/2 gives liminf y>=2/3. With w=z/T, H<=a and e<=3 imply w'>=y-(a+3)w. Scalar comparison gives liminf w>=2/[3(a+3)]. Hence U>=3a/2 forces T to zero, including equality; otherwise limsup T<1/b and limsup VT<=(3a/2-U)(a+3). These are eventual bounds; the proof does not substitute an immune nullcline into a transient trajectory. 3. Tail damping: -5/21/4 and trace<15+a on that tail, which proves the claimed uniform margin there. No assertion that the B interval itself is invariant is needed. 4. Ua: a decreasing energy either has a finite lower bound or tends to negative infinity. In the first case bounded kinetic terms and Q tending to negative infinity at zero force T away from zero. In the second case the energy is eventually below its minimum on every region T>=epsilon, forcing T to zero. This dichotomy is valid; reversing the sign of the logarithmic potential would invalidate it, but the sign here is correct. 6. U=a: the zero constant term of F makes Q'= -eF/T smooth at zero; my exact limit is V-Vcrit. The energy identity therefore extends to a compact tail closure that includes T=0. On its zero-dissipation invariant subset, p=r=0 implies F=0 for positive T, or y=1,z=0 at zero. There are finitely many equilibria, and connectedness of omega-limit sets reduces convergence to the equilibrium set to convergence to one equilibrium. Hyperbolicity is unnecessary here. 7. Local and fold classification: full cubic inequalities make every simple positive equilibrium a sink or an index-one saddle according to F', with no Hopf bifurcation. The physical fold is exactly the positive-V portion found above; the multiplicity bound rules out a higher-order interior fold. F_U is nonzero. The monotone curve and the maximum of f*n determine the stated two-root/zero-root sides; no projection-based stability inference is used. 8. Complete trajectories and basins: this is the strongest potential objection. An eventual energy inequality by itself would not exclude a saddle homoclinic. However, on any bounded complete positive trajectory, the y and w comparisons can start at time -s: their nonnegative initial terms can be dropped, and s can tend to infinity. The logistic comparison can be pulled back using (1/T)'>=-A/T+D with A=a-2U/3 and D=ab+2V/[3(a+3)]. Thus the comparison bounds, and damping, hold at every time on a saddle unstable branch. A homoclinic would have equal finite endpoint energies but strictly negative energy change, impossible. Global convergence leaves only the two sinks as branch destinations. Openness of the basins and the local hyperbolic exit map force the branches to reach different sinks, since both basins accumulate at the saddle. Pullback along the flow identifies the entire interior stable manifold as their common boundary. No unproved global graph representation is required. Controls and limits `control_runs.py` reconstructs the printed ODE independently, in log T and scaled effector coordinates. At U=0.6,V=2e-9 it finds positive equilibria T=257.90864268597187 and 9720.510428032607; full Jacobian eigenvalues give one unstable direction at the first and a sink at the second. Initial T=100 goes to the boundary sink and T=100000 goes to the upper sink. This reproduces ordinary bistability; it is not itself the added result. Ten diagnostics include very large immune initial conditions, U=a, U=3a/2, and couplings far outside the printed ranges. No simulation is used to establish absence of cycles. An extreme U