Skip to content
Vext Labs publishes "Beyond the Scaling Ceiling." Read the paper
Vext Labs

The proof surface

What we can show you.

Everything on this page is a thing you can open without asking us. Six papers, deposited. Three packages, on npm. One program award, with its scope caveat attached. The labels we use, and what each one means.

Papers

6 of 6, deposited 2026-07-27

All six papers are deposited on Zenodo under CC-BY-4.0, community zenodo.org/communities/vext-labs. Manuscript bodies stay withheld on this site while their claims are reconciled with authoritative evidence, the DOI is the door. Checked live against doi.org on 2026-09-03: all six resolve.

IDPaperScopeDOI
VL-2026-01 Beyond the Scaling Ceiling workshop draft · PROVEN-toy · BUILT/VERIFIED-CPU · not real-LLM scale 10.5281/zenodo.21628058
VL-2026-02 The Accretion Model BUILT/VERIFIED-CPU · PROVEN-toy · systems and trust, not a capability crown 10.5281/zenodo.21628429
VL-2026-03 The Leakage Signature reproducible artifact · toy + bounded real-LLM addendum (1.5B to 32B) · no frontier parity claim 10.5281/zenodo.21628524
VL-2026-04 Verifier-Centric Capability Growth Part IV is a proposed research program and reports no results 10.5281/zenodo.21628544
VL-2026-05 Sound Compounding all positive results are CPU-scale kill-tests; the real-LLM experiment is pre-registered and unrun 10.5281/zenodo.21628552
VL-2026-06 Capability Injection via Reverse Abliteration method paper; no production deployment claim is made 10.5281/zenodo.21628566

Body withheld on this site pending evidence reconciliation. Read the publication method and status →

Packages

npm registry, checked live 2026-09-03
PackageLicence / versionInstallStatus
@vextlabs/theron-agent-sdk

A published TypeScript package for building against Theron and JUWEL agent interfaces.

Inspect the package on npm

MIT / v0.4.1 npm i @vextlabs/theron-agent-sdk PUBLIC PACKAGE
@vextlabs/theron-cli

A published command-line client for JUWEL. Service access and configuration can depend on the account and endpoint in use.

Inspect the package on npm

Apache-2.0 / v0.6.1 npm i -g @vextlabs/theron-cli PUBLIC PACKAGE
@vextlabs/stoa-verifier

A published artifact intended to verify STOA receipts offline. This page does not claim a complete source-and-test rebuild from the current public repository, or that every JUWEL action already produces a stranger-rerunnable receipt.

Inspect the package on npm

Apache-2.0 / v0.1.1 npx @vextlabs/stoa-verifier ARTIFACT BUILT / CLAIM IN PROGRESS

Versions and licences re-checked directly against the npm registry on 2026-09-03; all three matched the versions previously recorded here. Use the npm listing as the current source for release changes.

Program support

announced 21 July 2026
ProgramAwardDateScope
Cloudflare for Startups Startup credits valued at $100,000 21 Jul 2026 This is program support, not investment, revenue, or money raised. Full announcement →

The labels

AGENTS.md, canon label set

Every claim we publish wears one of these seven labels. Including the ones we failed.

PROVEN

Measured and confirmed to hold, with the evidence attached.

BUILT

The thing exists and runs, checked against real output.

REGISTRY-ONLY

Recorded in a registry or a config, not yet exercised end to end.

GPU-GATED

The next step needs GPU compute that has not been funded or run yet.

NOT-BUILT

Described or planned, and nothing has been built yet.

UNAVAILABLE

Not accessible right now, whatever its underlying build state is.

KILLED

Tested and found false or unsupported, and kept in the record rather than deleted.

The section that makes the rest believable

What we do not have.

  • No matched real-model lifecycle-cost advantage exists.
  • There is no Vext AGI, ASI, all-domain, or generally self-teaching result.
  • The strong thesis "LLMs are not intelligence; the Accretion Model is" is KILLED. The defensible survivor is a receipted, non-forgetting verified-growth process, not intelligence, we claim the property, never the crown.
  • The one real-LLM run of finite_basis_meter (1.5B) is KILL-unbounded: not order-robust, and a clean re-run is owed before any final call.
  • An own-weights flagship model is NOT-BUILT / GPU-gated.
  • No certifications of any kind, no SOC 2, HIPAA, GDPR, ISO, FedRAMP, or BAA, anywhere.

Continue

Read the method behind the papers.

The research page explains the six gates, the publication contract, and today's withholding notice.