Datasets:

SZLHOLDINGS
/

thesis-v18-formal-verification

We present the Ouroboros Substrate (SZL Holdings, versions 14 through 18), a formally verified execution environment for agentic AI in which every governance invariant is machine-checked by the Lean 4 kernel against Mathlib v4.13.0. The central mathematical object is the Lambda-axis score, a nine-dimensional product-of-receipts governance scalar bounded by axiom system A1–A15. Axioms A1–A4 (monotonicity, positive homogeneity, Egyptian-fraction normalisation, max bound) characterise Lambda uniquely, proved in Lutar/Axioms.lean and Lutar/Bound.lean (DOI 10.5281/zenodo.20434308). Axiom A15 (SHA-256 collision resistance) is a cryptographic open problem under NIST FIPS 180-4, not derivable from Mathlib, and is labelled as such.

The v18.0 thesis (206 pages, DOI 10.5281/zenodo.20434276) contains 72 named theorem environments and 4 corollaries across eight chapters. Of the 31 theorems in Chapter 2 and 16 in Chapter 6, the Lean Czar audit (Chapter 7) certifies: 17 lake-verified with no sorry and no new axiom; 22 skeleton-pending with honest sorry markers; 1 open-problem-axiomatic (SBOM Lambda-chain total order, conditional on A15); and 1 pure metatheorem (Lean kernel soundness). The TH-V18 catalogue (16 stubs, TH-V18-01 through TH-V18-16) builds clean under lake build Lutar.Thesis with 13 axiom declarations and 8 in-code sorry tactics, each enumerable by grep at the chapter-bind tag 2026-05-28 (DOI 10.5281/zenodo.20434308).

The 32-module runtime executes 934+ inline assertions (exit code 0) and emits SCITT-compatible, SHA-256-chained receipts at every agent action. An OpenTelemetry SEMCONV extension surfaces Lambda scores as first-class indexed telemetry. Domain grafts span seven observability vendors (Splunk, Datadog, Dynatrace, New Relic, Better Stack, Honeycomb, Grafana), four cybersecurity platforms (Palantir, Palo Alto Networks, CrowdStrike, Fortinet), graph neural networks (PyTorch Geometric v2.7), sparse self-attention (rasbt DSA), agentic IDE governance (Cursor + Claude Opus 4.8), and sovereign-AI provenance (IQT Labs). Governance is enforced by Doctrine v6: a machine-checked ban on marketing superlatives applied at CI time.

All code, proofs, and data publicly archived via Zenodo DOIs 10.5281/zenodo.19944926, 10.5281/zenodo.20424992, 10.5281/zenodo.20424995, 10.5281/zenodo.20424996, 10.5281/zenodo.20431181, 10.5281/zenodo.20434276, 10.5281/zenodo.20434308.

Keywords: agentic AI governance, Lambda-axis calculus, Lean 4, Mathlib v4.13.0, dual-witness receipts, Schur concavity, PAC-Bayes bounds, Doctrine v6, SCITT receipt chain, sovereign AI, NIST FIPS 180-4, OpenTelemetry.