NOTICE

Eikos — NOTICE

This model is a fine-tune of a Qwen base model (Qwen3.5-4B or Qwen3.8-27B, by the Qwen team, Alibaba Cloud),
licensed under the Apache License 2.0. A copy of the Apache License 2.0 is included (LICENSE-Qwen). The base
model's weights were modified by LoRA fine-tuning and merged; the merged weights keep the original architecture
and checkpoint layout. Our own contributions (training code, data generators, evaluation code, serving code and
the fine-tuning deltas) are released under the MIT License (LICENSE).

Training signal
- Teacher labels: option probabilities produced blindly by GLM-5.3-Flash (MIT license, Z.AI; maximum reasoning
  effort), run by the authors on their own infrastructure.
- Synthetic decision items were written by Qwen3.8-27B (Apache-2.0) and GLM-5.3-Flash. Each item came with its
  gold answer and a short rationale, and was then labeled blindly by the teacher. We kept only the items where
  the teacher agreed with the gold answer. Rationales were used only as an auxiliary training loss; the released
  model does not generate text at decision time.
- EN<->PT translations of training items (view-consistency training, 4B only) were produced with Qwen3.8-27B.
- GSM8K answer-verification items use candidate solutions sampled from Qwen3.5-0.8B (Apache-2.0).
- Programmatic items (probability, temporal arithmetic, finance arithmetic, trading and trade-finance rules,
  compositional rules and rulebooks, long-context dossiers) were generated by our own code with exact answers.

Third-party data used for training (with attribution)
- FinEntity — ODC-BY 1.0 — Tang, Y. et al. (2023). FinEntity: Entity-level Sentiment Classification for
  Financial Texts. https://github.com/yixuantt/FinEntity
- TAT-QA — CC BY 4.0 — Zhu, F. et al. (2021). TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and
  Textual Content in Finance. https://github.com/NExTplusplus/TAT-QA  (used to build answer-verification items)
- GSM8K (train split) — MIT — Cobbe, K. et al. (2021). Training Verifiers to Solve Math Word Problems.
  https://github.com/openai/grade-school-math  (used to build answer-verification items)

Prompt format
- The structured decision prompt (evidence / criterion / lettered options) follows SemIf (MIT),
  https://github.com/theoleecj/semif — credit to its authors.

Trademarks and rule texts
- Incoterms® is a registered trademark of the International Chamber of Commerce (ICC). "UCP 600" refers to the
  ICC Uniform Customs and Practice for Documentary Credits. This project is not affiliated with or endorsed by
  the ICC. No ICC text was used; trade-finance rules were written in our own words for synthetic exercises.
- Trading rules (e.g., pattern day trading, wash sales, margin maintenance) were paraphrased in our own words
  for synthetic exercises. No FINRA or regulator text was used. The model does not provide legal, tax or
  investment advice.

Evaluation-only data (NOT used for training)
- JevBench (public items), DecisionBench, CUAD, FinQA, Twitter Financial News sentiment, WCB central-bank stance
  data, FinDVer, RuleArena, and the general battery tasks (MMLU-Pro, RewardBench, ASSIN2, BoolQ, Banking77,
  CLINC150, LegalBench, XNLI-es, GSM8K test) were used only to evaluate the model.