Beyond the Scaling Ceiling
Verified, non-forgetting, provable capability accretion with separate PROVEN-toy, BUILT-CPU, and GPU-gated claims. Not a real-LLM product claim.
Research archive
The public record of a claim.
Papers, artifacts, and boundaries for verified capability accretion. Every record says what was measured, what was built, and what remains proposed.
PROVEN-toy
Results measured on deliberately small curricula.
BUILT-CPU
Systems and artifacts that can be inspected today.
Proposed
Mechanisms and directions that are not results.
Published papers
Six records, with the scope beside the result.
Raw Markdown remains the canonical source artifact. Repository links point to the underlying paper or reproducibility material.
The Accretion Model
A growable, receipted model type with frozen-additive growth history. The contribution is systems and trust on CPU toys, not a capability crown.
The Leakage Signature
A falsification ledger and solver-certified enlargement with leakage audit. It includes toy work and bounded real-LLM demonstrations without a frontier-parity claim.
Verifier-Centric Capability Growth
A measured map of capability injection and negative results. The proposed portfolio is marked as proposed rather than presented as a completed result.
Sound Compounding Ratchet
A verifier-frontier ratchet and believed-versus-true test for self-improvement leakage. Positive results are toy-scoped and include CPU kill-tests.
Capability Injection via Reverse Abliteration
An early CIP method paper. It is research direction, not a claim that production JUWEL already rewires weights without forgetting.
Publication contract
A result earns a place in the archive by meeting the boundary.
State the wall
Before claiming growth, define what the base cannot solve under fair sampling.
Use a distinct checker
Admissions come from verifiers, tools, or solvers, not from same-model self-grading.
Preserve the prior work
Previously working parts stay byte-frozen or remain behind a regression gate.
Keep the failure
Killed mechanisms stay in the record because the negatives make the positive boundary meaningful.
Open tools
Check the record without asking us to be the referee.
The agent SDK, CLI, and offline receipt verifier are published for builders and for anyone who wants to inspect the work.