Deterministic structural verification, built on cellular sheaf cohomology. These papers cover the core incremental algorithm, its stability theory, and the applied verification systems built on it: one for evolving agent state, two for protein structures.
The core algorithm behind SIGMA. Shows how to maintain structural verification in constant time per edit, no matter how large the graph gets.
Classical computation of first sheaf cohomology H^1 on a cellular complex costs O(n^3) per edit through coboundary factorization, so a stream of edits costs O(m n^3). Under a bounded local geometry assumption (bounded cell size, stalk dimension, and nerve degree), each vertex insertion, edge insertion, or restriction map update touches only a bounded set of local coboundary blocks. The result is amortized edit processing that is constant in n, with deferred exact reconciliation at query time.
Applies SIGMA to AI agents that modify their own state. Proves that structural coherence can be maintained as agents evolve, with checks running in constant time per edit.
Update-time verification for self-evolving agents. Agent state is modeled as a growing graph equipped with a cellular sheaf, and coboundary checks, H^1 obstruction, Dirichlet energy, and Laplacian spectra detect structural contradictions before commit. Under explicit policy and regularity assumptions, the paper proves conditional Lyapunov stability for density and spectral-gap dynamics, and a cellular decomposition localizes recomputation to O(1) amortized per edit, sustaining maintenance at 5,000,000 vertices.
Proves that SIGMA's verification dynamics are stable: the structural properties the engine monitors converge to predictable equilibria rather than drifting or oscillating.
Conditional Lyapunov stability for a sheaf-governed graph whose topology grows under cohomological feedback. The edge-vertex ratio converges to a system equilibrium near 3.0, and the sheaf-Laplacian spectral gap converges to a positive equilibrium of 0.530. Validated on a 500-cycle run with ADF-confirmed stationarity, multi-seed (cross-seed CV 1.9 percent), and at scale via cellular decomposition: constant per-vertex eigensolve cost from 50,000 to 1,000,000 vertices, plus a streaming update path measured at 35 microseconds per edit.
Applies the same SIGMA verification engine to protein structures. Given a predicted structure and a native reference, it checks whether the prediction is structurally coherent and localizes failures to specific residues.
SATYA Protein builds a constant cellular sheaf over the residue constraint graph, where six-dimensional stalks encode per-residue spatial displacement and signed backbone dihedral deviation. The binary SAFE or UNSAFE verdict is gated on three independent checks: contact-network preservation (native contact fraction Q at least 0.90), localized damage, and chiral integrity. Failures are localized to specific residues, with no learned parameters in the verification loop.
Accepted at the 19th Computational Structural Bioinformatics Workshop (CSBW 2026), IEEE BIBM 2026, Dallas. A preregistered, computationally blinded test: three structural plausibility gates, SATYA's own included, advanced all 50 generated complexes, and the cohort still failed the mutation-site test stated before any candidate was scored.
Boltz-2 generated 25 wild-type and 25 M49I complexes of SARS-CoV-2 main protease with ensitrelvir. PoseBusters, MolProbity and SATYA's reference-free check advanced all 50; generator confidence advanced 25, of which 24 were wild type. No pocket-aligned ligand RMSD exceeded 1.01 Å, yet at the site where the minimum residue-49-to-ligand distance is 1.29 Å shorter in the mutant reference, the generated condition medians differed by 0.03 Å, and every wild-type candidate placed residue 49 closer to the ligand than the wild-type crystal structure does. Across the agreement descriptors recorded at freeze time, agreement was granularity dependent, and the preregistered residue-level endpoint was the most favorable of them. The frozen reference-free contrast rule declares a sampling floor and returned no classification at five seeds per condition; on KRAS G12D versus G12D/Y96S with MRTX1133 it returned RESOLVED at thirty seeds, and a post hoc analysis reversed the sign and found the response broad across the pocket. No construct-matched reference exists for the G12D arm. The paper keeps per-structure plausibility, cross-condition response and reference agreement apart, and reports what each one could and could not establish.
Every MCL1 protein-ligand structure in the PDB with a small-molecule component and published OpenFold3 predictions, run through the same SATYA route under a protocol pre-registered on Zenodo before any Boltz-2 model was generated. Boltz-2 models generated under that protocol are added as they complete.
A reading copy of the collection's records: what it contains, the state of each set, the conversion control, every protocol version, and the deposited file list, with each source marked as deposited on Zenodo or not yet deposited. No statement is made about the correctness of any structure, generator or verifier, and no frequency or cross-generator statement is made. The deposited files on Zenodo govern.
All benchmark artifacts are committed to the Locus transparency log with signed correctness receipts. Questions or collaboration: jason@invariant.pro.