Independent researcher · AI safety · based in Italy

I work on the algebraic limits of LLM-compiled symbolic reasoning and on scaling laws for external substrate verifiers. Two papers shipped in April 2026:

• Paper 1 — Structural Separation Theorems for Finite-Group Representations — a universal impossibility result for additive representations of finite groups, plus constructive torus / Peter-Weyl embeddings and a capacity bound K(N, ε) ≥ (π/ε)^N validated for N = 2..5.
• Paper 2 — Calibration Windows of Toroidal HRR Substrates — a three-regime scaling map, the margin formula m(K) ≈ 1 − C(V)√(KV/D), an operational design rule D*, an LLM-substrate crossover L*(K,V,D), and a controlled RLVR-vs-SFT null result at 0.5B parameter scale. (arXiv preprint pending endorsement.)

• Resolving SGD learnability of cyclic-group representations (Paper 2 OP1).
• Empirical test of the L* ∈ [3, 7]B crossover prediction.
• A third paper on hybrid architectures — when to route to the substrate verifier vs let the LLM self-check, using m(K) as the operational criterion.

Explicit hypothesis pre-registration, numerical validation, and retraction discipline when signals do not replicate. The earlier AGI-embryo framing of this research program was retracted after five separate intrinsic-fitness-signal failures in a pre-registered test sequence.

• Email: daniel.culotta@gmail.com
• LinkedIn: daniel-culotta-ml
• Zenodo: Paper 1 DOI 10.5281/zenodo.19642604

Open to collaboration, technical discussion, and reviewer feedback on either preprint.