NOTICE — third-party software & provenance
===========================================

Crosslight is released under the license in the LICENSE file. This NOTICE documents
its provenance and its relationship to third-party software, content, and models.

No bundled third-party content or model weights
-----------------------------------------------
Crosslight ships with an EMPTY document index. It does not include, redistribute, or
embed any books, articles, datasets, copyrighted text, or model weights. The
library ([B#]) tier is a capability you fill with documents you own or are licensed
to use. Language models are downloaded separately by the user via Ollama and are
governed by their own licenses; Crosslight redistributes none of them.

Third-party Python dependencies
-------------------------------
All runtime dependencies are OPTIONAL (the core retrieve -> rerank -> verify ->
synthesize loop runs on the Python standard library) and are permissively licensed.
They are not bundled; users install them via pip, under their own licenses:

  - numpy               BSD-3-Clause
  - beautifulsoup4      MIT
  - charset-normalizer  MIT
  - argostranslate      MIT

External services
-----------------
The web and scholarly tiers call public third-party services at runtime, on the
user's own machine (e.g. OpenAlex, Semantic Scholar, arXiv, and a configurable web
search backend). Crosslight redistributes none of their data; such use is subject to
each service's own terms.

Techniques
----------
Crosslight implements well-known, publicly documented techniques — retrieval-augmented
generation; BM25 + embedding hybrid retrieval with reciprocal-rank fusion;
LLM-as-judge reranking; claim/source consensus ("disagreement") checking; and
standard indirect-prompt-injection defenses. Methods and algorithms are not subject
to copyright. No third-party source code or text was copied into this project.

Authorship
----------
Crosslight was engineered with substantial AI assistance — the bulk of the build and
review was performed by AI models (Anthropic's Claude as lead engineer, with review
and critique from other AI models) under human direction. It is released as a
standalone, original work.
