Stratified Sheaf–Topos Autopoiesis

Can a neural model improve on its parent for every label law consistent with the evidence? Stratified Sheaf–Topos Autopoiesis gives an exact answer for scalar ReLU probability networks under a uniform input law. A strictly improving successor exists precisely when the parent’s prediction asymmetry exceeds the allowed label asymmetry. If the parent falls within that budget, its own predictions define an admissible label law that makes a universally improving successor impossible. The paper turns this criterion into a construction. It produces rational neural successors and verifies their worst-case Brier loss against a bound that applies to every possible successor. The largest certified gap to the minimax optimum was below 2.39 × 10⁻⁹ Brier units across 264 replayed comparisons. In 12 controlled trajectories, all models completed four consecutive certified improvements under progressively tighter synthetic budgets, with no retraining or new labels in that experiment. A separate paired-label audit shows how fresh independent observations can support an asymmetry bound. Together, these results give a testable account of what verified self-improvement can achieve and exactly where the task assumptions force it to stop.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24234244
Primary Topic
Neural Networks and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Stratified Sheaf–Topos Autopoiesis

Ved Prajapati
Zenodo (CERN European Organization for Nuclear Research)
Neural Networks and Applications
preprint

Stratified Sheaf–Topos Autopoiesis

Ved Prajapati
preprint en

Abstract

Can a neural model improve on its parent for every label law consistent with the evidence? Stratified Sheaf–Topos Autopoiesis gives an exact answer for scalar ReLU probability networks under a uniform input law. A strictly improving successor exists precisely when the parent’s prediction asymmetry exceeds the allowed label asymmetry. If the parent falls within that budget, its own predictions define an admissible label law that makes a universally improving successor impossible. The paper turns this criterion into a construction. It produces rational neural successors and verifies their worst-case Brier loss against a bound that applies to every possible successor. The largest certified gap to the minimax optimum was below 2.39 × 10⁻⁹ Brier units across 264 replayed comparisons. In 12 controlled trajectories, all models completed four consecutive certified improvements under progressively tighter synthetic budgets, with no retraining or new labels in that experiment. A separate paired-label audit shows how fresh independent observations can support an asymmetry bound. Together, these results give a testable account of what verified self-improvement can achieve and exactly where the task assumptions force it to stop.

Zenodo (CERN European Organization for Nuclear Research)
Neural Networks and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon