Σ-Model — Schema-Coherence Suppression & Compositional Generalisation Failure

A dynamical-systems framework for compositional generalisation failure. Core idea: standard gradient-based training drives agents into a stable low-schema-coherence equilibrium (the σ-trap) — high in-distribution accuracy coexists with systematic out-of-distribution compositional failure because depth accumulates while schema coherence is suppressed.

The active deliverable is a single narrow manuscript (the σ-Trap paper, targeting TMLR after the JAIR desk-rejection) plus an arXiv companion technical report preserving the extended framework (proxy architecture, cognitive extensions, multimodal coverage, benchmark protocol).

---

Repository Structure

Path	Contents
paper/	σ-Trap manuscript (TMLR format) + arXiv companion (paper/companion/) + planning docs (paper/planning/)
code/sigma_align/	Reusable Python package (ODE, config, utils, monitoring, evaluation)
archive/	Read-only historical artifacts: thesis monograph (archive/thesis/), decision docs (archive/Σ-Align/), old code, datasets, experiment results. Do not modify.
docs/adrs/	Architecture Decision Records
hbar_env/	Python virtual environment

Current Status

• Paper 06 (Σ-Model) was desk-rejected by JAIR (2026-07-15) on exposition/notation, overbroad claims, and scope grounds. Pivoting: narrow σ-Trap paper → arXiv → TMLR, experiment-gated. See paper/planning/roadmap.md for the phase plan.
• The thesis monograph (10 chapters) and AGI-safety publication pipeline are archived (reversible via git) — the project is re-centered on the dynamical-systems account of compositional generalisation failure.

Build

Task	Command
Build the paper PDF	make paper06 (from root) or make pdf (from paper/)
Lint	ruff check code/sigma_align/
Run evaluation	PYTHONPATH=code:$PYTHONPATH python -m sigma_align.monitoring.evaluation
CLI	sigma-evaluate (installed via pip install -e .)

Activate the venv first: source hbar_env/bin/activate.

Citation

@misc{sigma-model2026,
  title={The {$\Sigma$}-Model: Schema-Coherence Suppression as a Dynamical
         Mechanism of Compositional Generalisation Failure},
  author={{Basyirin Amsyar Basri}},
  year={2026}
}

License

MIT