Skip to content
Allagma

Licenses and source lineage

Allagma’s core implementation is original code and prose under the repository’s MIT license. The Python interpreter is a prerequisite, not redistributed here. The default runtime has no third-party Python dependencies. No ARIS, RRSI or AntOmniEvo source code or skill text is vendored.

The adopted design draws on ARIS (MIT), RRSI (Apache-2.0), AntOmniEvo (Apache-2.0), Agent Skills, Copier, uv and nf-core. These are design references, not installed runtime dependencies or claims of integration. Pinned mechanisms and local adaptations are listed in design lineage. The license descriptions are the findings of the supplied design audit, not a new legal review of every upstream file. Any future source reuse must retain the applicable file-level license and notices.

The planning specifications in docs/specification/ were supplied by the repository owner and copied with a digest provenance record. External links inside them identify their own sources.

Native Codex weight EMA diffusion study

Section titled “Native Codex weight EMA diffusion study”

studies/ema-2d-diffusion/reference/ retains selected source files from SakanaAI/AI-Scientist at commit 1de1dbc1f4ee2c5f61e9c94348d55eb51d7fa2eb, with a retrieval and digest record. The reference files and the adaptations in domain/science.py and domain/runner.py are subject to the AI Scientist Source Code License, version 1.0. This source is outside the core’s default runtime. The license accompanies frozen scientific materials.

REFERENCE.md identifies the adapted mechanisms and modifications. The study’s machine-generation disclosure applies to its code, analyses, manuscript drafts and reports; the final published manuscript repeats that disclosure prominently. Python packages are installed only into an ignored study-owned environment; their versions are pinned in requirements.lock and their package licenses apply to those installations.

v0.3 study development sources

Section titled “v0.3 study development sources”

studies/ema-schedule/development/science.py copies the preceding study’s adaptation; the same license is retained at studies/ema-schedule/reference/LICENSE. Its follow-up runner also uses those adapted mechanisms. The follow-up disclosure identifies the source and AI-generated material.

studies/core-culp/reference/ retains CORE-Bench scoring/prompt files from siegelz/core-bench commit e32a2980e72fe6eb04ee04eb749458f570625663 under its included MIT license and the original capsule-6460826 archive. Capsule code is accompanied by its MIT license; dataset files retain their CC0 1.0 dedication. The provenance record identifies the exact archive. These are study/evaluation resources, not dependencies of the default conformance kit. Original reference answers and scorer code remain outside evaluated agents’ writable/readable workspaces.

The modular-addition pilot is an original small MLP implementation inspired by Power et al., with the adaptation stated in its study README. No upstream transformer implementation is copied into that pilot.

The frozen v0.3 modular-addition evaluation additionally supplies selected reference files from openai/grok at commit 3d64b1d8c1d595dd8ebdb7771998823f1b14c7b3, with the OpenAI copyright notice and MIT license, inside candidate packages’ inputs/materials/openai-grok-reference/. The freeze manifest pins those bytes. They are reference materials for a bounded MLP study, not evidence that the original transformer experiments or training horizons were reproduced.

Documentation and launch media

Section titled “Documentation and launch media”

The separate site/ project uses Astro Starlight and its locked Node development dependencies. Those dependencies retain their upstream package licenses and do not enter the default Python runtime. The site build includes their shipped copyright and license notices in generated/third-party-licenses.txt, linked from the page metadata. That appendix includes build-time dependencies as well as browser code; it does not relicense Allagma or the scientific archives. The workflow graphics, mark and social preview are original Allagma assets; uv and marimo supplied editorial inspiration, with no copied branding or graphics.

The flagship film captures actual local software. Narration is synthesized locally with Kokoro-82M v1.0’s af_heart voice (Apache-2.0 model) and kokoro-onnx (MIT engine). The model and rendering dependencies are not redistributed in the film or core. The narration notices retain pinned source links, attributions and a full Apache license copy. Earlier private Git history retains superseded Apple-voice renders; the publication review records that separate history issue. The film’s final chapter and media/evidence/ema-r07.png reuse the machine-generated EMA figure produced with AI Scientist-adapted code. The full study license accompanies that figure and the video distribution, and the video, transcript and site disclose the reuse and machine generation. The compact core source archive excludes third-party study adaptations and scientific dependency wheels. See artifact distribution.

Canonical source · Generated from the maintained repository document at this build’s commit.