Beyond the Scaling Ceiling: Verified, Non-Forgetting, Provable Capability Accretion

A study of how an AI model could grow new skills in checked steps. Each result has a tag: PROVEN-toy (tiny model), BUILT/VERIFIED-CPU (working code) or GPU-GATED (waiting on big chips).

Annalea Layton · Vext Labs · 2026-07-02 · 17 pp · ≈25 min read

workshop draft
PROVEN-toy
BUILT/VERIFIED-CPU
not real-LLM scale for CIP forgetting product claims

PDF · 17 pp
DOI 10.5281/zenodo.21628058
Code · coming soon
Data · coming soon

Cover · A pixel pattern made from this paper’s ID, “beyond-scaling-ceiling” · Art, not a chart

On this page · 4 sections

How we state results
Abstract
Figure 1
What we do not claim

Result as of 2026-07-02 · Tags: PROVEN-toy + BUILT/VERIFIED-CPU

The raw data for Figure 1 is in ledger/juwel_cip_pipeline_proof.json. It has no public link yet. Until it does, the figure names the file it came from.

How we state results

This paper is an honest answer to the scaling-law papers.¹ Each result carries the tag our master list allows. Nothing tagged toy or GPU-gated is claimed for full-size AI models.

Abstract

The scaling era’s implicit promise was that scaling next-token prediction on a frozen model would keep buying capability. We report the honest inverse: the data-scaling exponent is architecture-invariant, so more scale moves the level of the curve, never its slope.

We introduce the Accretion Model, a model type that grows only by frozen-additive, checker-gated, receipted increments, and whose growth history a stranger can verify offline. Its contribution is architecture, systems and trust, not capability.

Figure 1

Two charts from the paper. After growth, the score on the new topic falls from 4.51 to 3.48 (domain NLL, lower is better). The general score moves from 21.3 to 20.1 (perplexity), inside the ×1.15 limit.

Figure 1 · CIP adds 4 layers to a small, frozen 0.5B model without breaking what it knew. Domain NLL 4.51 → 3.48. General perplexity 21.3 → 20.1 (limit ×1.15). Tag: PROVEN-toy. Source: ledger/juwel_cip_pipeline_proof.json · as of 2026-07-02.

What we do not claim

Each idea below failed our tests, and that counts as a result. We wrote down the pass mark before each test. We compared each one against a fair baseline.

KILLED: Capability from nativeness: a native in-forward-pass loop beats an external harness on capability
KILLED / known-result: A learned world-model brain discovers verified capability more efficiently than best-of-N
KILLED / rename-of-wrapper: Nativeness is structurally cheaper
KILLED / known-result: An internal competence signal recovers routing with no task label
NOT DEMONSTRATED · GPU-gated: The fuel bill is bounded by a small basis of execution classes
KILLED / built-in-answer: Weight-accretion is a provable moat over RAG
NEVER CLAIMED: “Smarter”, “more capable”, “human-level AI”, “first intelligence”

Footnotes

1. The tags come from the lab’s master list of results. The paper repeats them. It does not pick them. ↑ back

References

Kaplan et al. (2020). Scaling Laws for Neural Language Models.
Hoffmann et al. (2022). Training compute-optimal large language models (“Chinchilla”).
Rusu et al. (2016). Progressive Neural Networks.
Mallya & Lazebnik (2018). PackNet.
Serrà et al. (2018). Overcoming catastrophic forgetting with hard attention to the task.
Laurie et al. RFC 6962: Certificate Transparency.

Authors

Annalea Layton · Vext Labs, Inc. · ORCID 0009-0001-3695-9569

Acknowledgments

None stated in the source paper.

Cite

Layton, A. (2026). Beyond the Scaling Ceiling: Verified, Non-Forgetting, Provable Capability Accretion. Vext Labs. https://doi.org/10.5281/zenodo.21628058

@techreport{layton2026beyond,
  title  = {Beyond the Scaling Ceiling: Verified,
            Non-Forgetting, Provable Capability Accretion},
  author = {Layton, Annalea},
  institution = {Vext Labs},
  year   = {2026},
  type   = {Workshop draft},
  doi    = {10.5281/zenodo.21628058}
}

DOI 10.5281/zenodo.21628058 CC-BY-4.0