VectorCertain builds AI safety governance technology: mechanisms that verify whether an AI decision's evidence is independent, whether the decision may execute, and whether that governance can be proven to parties who never see the model. The work is organized as ten projects across three tracks. The research track measured cross-model failure correlation of 0.95 across 13 frontier models — against a 0.40–0.60 literature assumption — and then engineered effective independence from 1.007 to the theoretical maximum of 13.000. The patent track converts those measurements into filed claims, with the 77-claim HCF2-SG framework as the hub every other filing incorporates by reference. The platform track ships it: SecureAgent ACA, its certification and consortium layers, and a self-contained assessment report requiring nothing of the recipient. Each entry below links to its full technical page.
HOW TO READ THE DOCKETThree tracks, one dependency chain
The tracks are not parallel product lines; they are stages of the same argument. The research track establishes the problem empirically: if 13 frontier models fail with a mean correlation of 0.95, then any safety architecture built on multi-model agreement — redundancy, consensus, majority voting — delivers a fraction of the protection its design implies. The 28 shared vulnerability signatures identified above a 0.70 threshold locate exactly where that shared failure concentrates.
The patent track converts the measurement into enforceable mechanism. HCF2-SG (P-01) claims independence verification itself; MRM-CFS (P-02) claims execution gating at 3.6 milliseconds; Cyber-SG (P-08) and zkML-SG (P-09) carry the mechanism into security enforcement and cryptographic proof. All four share the December 2025 priority anchor, and the three application filings incorporate the 77-claim hub by reference — one framework, licensed once, surviving model and vendor replacement.
The platform track is the existence proof. SecureAgent ACA (P-05) shipped the stack to general availability through a 22-sprint engineering arc with zero failing tests across 36,181 regression executions; MYTHOS (P-06) keeps the governance layer adversarially certified against six attack families; ZGTID (P-07) extends assurance across institutions without exposing any single member; and the Tier A report (P-10) puts a working artifact in an evaluator's hands with nothing to install. Research measured it, patents claimed it, the platform runs it.
The docket is additive: new filings, studies, and releases append as numbered entries without restructuring this index.
CONTACTWalk the docket with us
Every figure on this page traces to sealed, hash-verified validation artifacts — including the measurements that contradicted our own published estimates, which we recorded rather than reconciled away. Technical briefings are available for enterprises, evaluators, and standards bodies.
Request the technical briefing