# Public release review, version 0.1.1

## Authorization

Guilherme Monteiro, project owner and maintainer, authorized public visibility
for the three repositories listed below on 2026-08-26. The authorization was
given directly in the controlled project session used with the authenticated
`guicybercode` Hugging Face account.

| Repository | Artifact scope |
|---|---|
| `guicybercode/br-sovereign-llm-research` | original protocol, code, documentation, configuration, tests, and local engineering evidence |
| `guicybercode/br-sovereign-llm-corpus` | eight project-generated, AI-assisted synthetic pt-BR validation documents and their derived Parquet/audit artifacts |
| `guicybercode/br-sovereign-llm-smoke` | a 149,696-parameter engineering checkpoint trained for six CPU steps only on those eight documents |

This approval does not cover a future scientific corpus, target-scale weights,
optimizer state, raw logs, cluster configuration, or any artifact not present
in the content-addressed public manifests.

## Rights review

- Research code and prose in this release were developed for the project with
  AI-assisted drafting and are released under Apache-2.0 by the project owner
  to the extent of his rights. Bibliographic references and external URLs do
  not bundle third-party papers or datasets.
- Every distributed corpus row is a project-generated, AI-assisted synthetic
  fixture. The owner reviewed and adopted the eight texts for this release and
  selected Apache-2.0 to the extent of his rights. No Wikimedia, OBMEP,
  official-act, ProEdu, SciELO, benchmark, web-crawl, or other third-party
  document is included.
- The smoke weights were derived only from the eight released fixtures. The
  owner selected Apache-2.0 for this smoke checkpoint. That decision does not
  preselect a license for future weights trained on another corpus.
- Dependencies are identified by the lockfile and SBOM but are not vendored or
  relicensed by this release.

This is a scoped release and rights inventory, not a representation that
external legal counsel issued an opinion or that AI-assisted text is unique.
A future release containing real corpus documents or scientific weights
requires a new item-level rights and legal review.

## Privacy and security review

- The corpus fixtures contain no observed personal data or third-party text.
- The only named person is the author/maintainer, who supplied the displayed
  name, FIAP affiliation, and ORCID for publication.
- Format-aware scans found no credentials, private URLs, local home paths,
  internal hostnames, private IP addresses, Slurm accounts, or site storage
  paths in the release candidates.
- Raw logs, optimizer state, and the internal `.pt` recovery checkpoint are
  excluded. The smoke repository distributes safetensors weights only.

## Scientific-claim review

The corpus and smoke model are engineering fixtures. They do not establish
language-model quality, target-scale recovery, Brazilian supercomputer use,
data residency, energy consumption, portability, or national sovereignty.
Those measurements remain `not_measured`.

## Maintenance and takedown

Guilherme Monteiro is the named maintainer. To request correction, attribution
change, or withdrawal, open a Community discussion in the affected Hugging
Face repository and identify the repository, commit, artifact path or document
ID, and requested action. Do not place sensitive personal data in the public
discussion; request a private follow-up when necessary.

A credible urgent report can cause the repository to be hidden while it is
reviewed. Confirmed issues are recorded against the affected content hashes.
A corrected artifact receives a new version and manifest; published history is
not silently rewritten.

## Decision

The three version 0.1.1 packages are approved for public visibility within the
scope above, subject to a final exact-inventory scan before the visibility
change and an unauthenticated byte-for-byte verification afterward.