Flagship thesisworkshop draft
Beyond the Scaling Ceiling
Verified, non-forgetting, provable capability accretion: the clearest synthesis of the lab thesis and the public framing for faster, cheaper model growth.
Vext Labs · research archive
Train models faster. Better. Cheaper than scratch.
Our research program is verified capability accretion: freeze what works, add new capacity only when a checker can certify it, and publish the growth trail so a stranger can verify what changed. The goal is not another expensive full retrain. The goal is a model that can grow without forgetting.
FasterAdd capability in rungs, not full training runs.
BetterGate new skills with external verifiers and abstain when uncertified.
CheaperKeep correct once, instead of re-paying context or scratch retraining.
If we tested it, we publish it. If it failed, it still counts. Scope labels stay attached.
Published nowdraft papers and artifacts
The core papers. Readable today.
These are public drafts, not polished claims decks. They preserve the important boundary: toy-scale means toy-scale, GPU-gated means not yet proven, and killed hypotheses stay in the record.
Model typebuilt CPU artifact
The Accretion Model
A growable, no-forgetting, externally fueled model type with a cryptographically verifiable growth history. The honest contribution is systems and trust, not a capability crown.
Artifactreproducible
The Leakage Signature
Falsification, solver-certified enlargement, and a logged artifact trail for separating real capability growth from leakage, elicitation, or built-in answers.
Mechanism mappreprint
Verifier-Centric Capability Growth
A measured map of capability injection, recurrent-depth boundaries, negative results, and proposed verifier-centered mechanisms.
Ratchet looptoy scope
Sound Compounding Ratchet
A verifier-frontier ratchet for capability acquisition, plus the believed-vs-true test for self-improvement leakage. Positive results are toy-scale.
CIP origindraft
Capability Injection via Reverse Abliteration
The early CIP framing: additive growth as the opposite of destructive full retraining, with regression gates and no-forgetting as the operating constraint.
Methodclaims discipline
How the archive stays honest.
01
Pre-register the wall
Before claiming growth, establish what the base cannot solve under fair sampling.
02
Use a different checker
Admissions come from external verifiers, tools, or solvers, not same-model self-grading.
03
Freeze old skill
Growth is additive. Previously working parts stay byte-frozen or regression-gated.
04
Publish the failures
Killed mechanisms are not footnotes. They are the boundary that makes the positives meaningful.
Full archiverelease queue
Everything else gets published through a gate.
The internal tree contains hundreds of notes, ledgers, experiment logs, council digests, and launch-adjacent strategy files. The publication plan is to release the research, not accidentally publish private ops. Each archive wave gets a claim label, redaction pass, source path, and stable public URL.
Wave 1: flagship papersPublished on this page now: synthesis, Accretion Model, leakage signature, verifier-centric growth, sound compounding, and reverse-abliteration CIP.
Wave 2: evidence ledgersPublish the research program ledger, research index, reproducibility commands, and claim-safe evidence tables after redacting internal memory refs and credentials.
Wave 3: mechanism dossiersRelease the council architecture, inference-cost, verifier-boundary, world-model, and continual-learning dossiers as dated technical notes.
Wave 4: experiment archiveExport raw results, JSON ledgers, scripts, and negative-result reports with reproducibility instructions and exact status tags.
External one-liner: Vext Labs is researching verified capability accretion: adding model capabilities faster and cheaper than retraining from scratch, while preserving old capabilities and proving each growth step.