RoGuard optional v9 model components

The RoGuard Python source is MIT licensed. Third-party model weights and model
derivatives retain the following separate terms. This notice describes candidate
components; it does not establish qualification or release status.

Romanian encoder
  dumitrescustefan/bert-base-romanian-cased-v1
  Revision: 37fb0ffb4bc4f7c4cde429626775685fb18f234f
  Copyright (c) 2020 Stefan Dumitrescu
  MIT license: licenses/ro-UPSTREAM-LICENSE
  https://huggingface.co/dumitrescustefan/bert-base-romanian-cased-v1

Ukrainian encoder
  youscan/ukr-roberta-base
  Revision: 8149bde480a6df9014aa934c3f8af30858dea5a6
  Apache License 2.0: licenses/Apache-2.0.txt
  https://huggingface.co/youscan/ukr-roberta-base

Multilingual entailment encoder and derived checkpoints
  MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7
  Revision: b5113eb38ab63efdd7f280f8c144ea8b13f978ce
  MIT license, copied unchanged from the pinned model repository:
    licenses/nli-UPSTREAM-LICENSE.txt
  https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7

Interpreter base and derived adapter
  Gemma 3 text 4B instruction model, MLX community 4-bit conversion
  Revision: 4f665a4c50ecfe4ecdc34056ab52fe3e3c4abf9e
  https://huggingface.co/mlx-community/gemma-3-text-4b-it-4bit
  Gemma is provided under and subject to the Gemma Terms of Use found at
  ai.google.dev/gemma/terms
  Copies: licenses/Gemma-Terms.txt and licenses/Gemma-Prohibited-Use-Policy.txt

Use and distribution of this Gemma 3 component and its model derivatives are subject to the
Gemma Terms of Use, including the use restrictions incorporated by Section 3.2.
Those restrictions are conditions of the optional model distribution and are
not replaced by the RoGuard source-code MIT license.

Gemma 4 base and derived shared-attention adapters
  Google Gemma 4 E4B instruction model
  Upstream revision: fee6332c1abaafb77f6f9624236c63aa2f1d0187
  MLX community 4-bit conversion revision: 475b9088d29754a3379866cf5aeb6b41acd313c2
  https://huggingface.co/mlx-community/gemma-4-e4b-it-4bit
  Google's pinned model card declares Apache-2.0 and links to:
    https://ai.google.dev/gemma/docs/gemma_4_license
  Copy: licenses/Apache-2.0.txt
  The conversion model card retains a license: gemma metadata tag. Both original
  model cards are preserved unchanged as licenses/gemma4-google-README.md and
  licenses/gemma4-mlx-README.md; URLs and hashes are in licenses/gemma4-sources.json.
  The upstream base files are unmodified. RoGuard separately trains low-rank
  query/output projections in the shared-attention layers and classification heads.

Modifications by RoGuard
  The Romanian and Ukrainian classifier checkpoints are fine-tuned derivatives
  of the listed encoders. The field heads and decision combiners were trained
  for RoGuard. The interpreter adapter contains RoGuard-trained low-rank
  parameters for the listed Gemma base; the downloaded upstream base files are
  unchanged. The entailment checkpoints are RoGuard fine-tuned derivatives of
  the listed multilingual model. Model-selection records, hashes and limitations are recorded in
  research/v9-stabilization and docs/V9_STABILIZATION.md.

No upstream author or model provider endorses RoGuard through this notice.
