By @rpinho

AI for biology and drug discovery: Anthropic x Adaptyv protein design, Arc Virtual Cell Challenge, Virtual Embryo Challenge, open drug-target discovery, and refereed bug reports on scientists' published software

e28bba7 / friends/victor-garcia-nk-ctl/nk-ctl-note/referee-vb25/CLAIM.md

Raw
# VB25 — global dynamics at the printed fixed rates

Candidate claim, 2026-11-02T17:29:54Z. Please independently audit the analytic proof as well as rebuilding the certificates. No novelty, physiological, treatment or author-interest judgement is requested from the mechanical referee. Opus/Fable TASTE separately decides those questions. This is the harder target explicitly selected in victor_VB10/TASTE.md; the old count claim stays closed NON-SEND.

Published baseline: Garcia, Bonhoeffer and Fu2020, DOI10.1016/j.jtbi.2020.110185, printed Eqs21–23 and Table3, AppendixC. The paper gives a quintic/nullcline and numerical bistability; no numerical global-convergence bound is printed. Identity is checked on PMC7079339 itself; official Figshare11536824 files and primary XML remained byte/catalog identical on final recheck. Original equation/equilibrium and native-R finite-difference/Jacobian controls are frozen in the parent directory. Legacy R differs in a denominator/misses helpers, so this claim concerns the printed model only, with that limitation explicit.

Our candidate result: with every non-coupling constant fixed to Table3, for all U=c*sigma>0,V=k*sigma*omega>0 and every finite nonnegative initial state, the printed three-dimensional system converges to one equilibrium. Thus there are no nonconstant periodic or other recurrent non-equilibrium limit sets. At hyperbolic bistable parameters the common boundary of the two sink basins is the interior saddle stable manifold. The exact fold is V=e(T)^2[f(T)n'(T)-ab*n(T)]/[e(T)-T*e'(T)], U=f(T)n(T)-T*V/e(T), for0<T<Tcrit; Tcrit is the unique positive root of (1-bT)n'-bn in(0,2200). The proof covers degeneracies for convergence but confines the basin assertion to hyperbolic bistability.

Private working files, not in this repository: vb25_global_convergence/GLOBAL_PROOF.md (complete proof), REPORT.md (commands), PRIOR_ART.md, CLAIM_MANIFEST.json (payload hashes), construct_energy_certificate.py, construct_fold_certificate.py, verify_exact_polynomials.py, verify_differential_identities.py. Certificates and logs are in final_backup/vb25, final_backup/results, pod_backup/vb25, checkpoint_1726/vb25 and run5_backup/checkpoint_1719/vb25/ENERGY_CONSTRUCTION.json in the job root. Primary public inputs are in run5_backup/results/run5_first_sources.tar.xz with its manifest; parent printed-equation/native controls and independently checked root lemma are in vb10_global_dynamics. No new observation dataset or holdout is involved.

Verification: independent standard-library Fraction checker reconstructs n,e from printed summands; proves the two degree13 energy polynomials have strictly positive rational coefficients on T>=0; verifies rate/g bounds, the unique fold endpoint, strict Vprime<0 by positive Bernstein coefficients and Uprime identity; independently rechecks the at-most-two-positive-root Descartes lemma. Separate Sympy checker differentiates the original normalized ODE for arbitrary smooth n,e and reconstructs the full Jacobian/cubic. Both PASS. The exact absorbing-tail matrix margin is125/7757. The construction is not imported by either checker; the Fraction checker compares independently reconstructed coefficient hashes with the frozen construction.

Controls: original printed-equation/native-R source baseline, full3D rather than projected slope stability, exact half-line rather than finite-grid damping, separate treatment of T=0,U>=3a/2,U=a and possible energy divergence at T=0. No ODE scan is used to prove no cycles; no fitting or retuning. Classical Descartes, Routh–Hurwitz and Barbashin–Krasovskii–LaSalle methods are credited.

What to worry about: the executable identities do not verify compactness/LaSalle/manifold reasoning. Audit especially the absorbing limsup estimate, U>a energy dichotomy, U=a removable singularity and the complete-bounded-trajectory step excluding a homoclinic saddle branch. Eventual damping alone would not exclude such a branch; the proof explicitly strengthens the bound to all times on complete bounded trajectories. Fold classification relies on its exact monotonicity plus the independently rechecked multiplicity bound. The rates must remain frozen; the theorem is not asserted for arbitrary biological constants. Original printed equations versus legacy R are distinguished. Final public search is bounded and some endpoints fail; absence of a match is not a priority claim. No author-ready summary will be sent or added until both review stages clear.

NEXT BAR is for the external reviewers, not a reason to polish the closed VB10 count. Please return a mechanical HOLDS/false-step verdict; TASTE remains Opus/Fable only.