NOTICE

Latence Framework
Copyright 2026 Dennis Dickmann (latence.ai)

Required Notice: Copyright 2026 Dennis Dickmann
(https://github.com/ddickmann/latence)

This product includes software developed by Dennis Dickmann (latence.ai).
Licensed under the Apache License, Version 2.0 — SPDX identifier Apache-2.0,
reproduced verbatim in the LICENSE file. Commercial use is granted by that
licence; no separate agreement is required.

PROVENANCE (not part of the licence terms): this repository carried an
Apache-2.0 LICENSE from its first commit until 2026-07-29, was relicensed to
PolyForm Noncommercial 1.0.0 on that date (ADR-0055), and was relicensed back to
Apache-2.0 on 2026-08-15 (ADR-0063, which supersedes ADR-0055). Neither change
reached anyone, because the software was never published under either: the
GitHub repository has been private since creation, no
distribution was ever published to PyPI, no release was ever cut, no version tag
exists, every commit is authored by the sole copyright holder, and no contributor
licence agreement has ever been signed. So no one holds a grant under either
superseded state, and Apache-2.0 is the only licence this software has ever been
distributed under. The reasoning is in
docs/adr/0063-relicence-to-apache-2-0.md; the LICENSE file itself is the
canonical Apache-2.0 text and nothing else, so that scanners match it byte for
byte.

Under Apache-2.0 this project's inbound and outbound terms are the same, and
commercial use — including use inside a for-profit company, in a product, or as
a hosted service — needs no permission and no agreement from the copyright
holder. COMMERCIAL.md exists only to say so to readers arriving from the
superseded PolyForm-era pages.

That licence covers the Latence Framework's own code and nothing else. The
third-party software, model weights and datasets listed below and in
THIRD-PARTY-LICENSES.md carry THEIR OWN licences, which the framework's licence
neither changes nor overrides.

================================================================================
Third-party model weights offered as DEFAULT reference Providers
================================================================================

The framework never bundles model weights. It names Capabilities and lets
Providers fulfil them; a Provider downloads its checkpoint only when a Pipeline
configures it. The models below are the permissively-licensed DEFAULTS the
reference Providers point at. Weights AND code licenses were verified separately
per model, from 2026-07-06 onward; each Provider's own ProviderProfile records the
citation and the verification date, and THIRD-PARTY-LICENSES.md carries the
full table (see
docs/adr/0012-permissive-only-reference-providers-gemma-opt-in.md).
All are Apache-2.0 or MIT:

  - lightonai/LightOnOCR-1B-1025 (Parse/OCR, `parser.lighton`)      — Apache-2.0
  - lightonai/LightOnOCR-2-1B (Parse/OCR, `parser.lighton_vllm`)    — Apache-2.0
  - fastino/gliner2-multi-v1 (Fused Entity+Relation)                — Apache-2.0
  - fastino/GLiNER2-Guardrails-PII-Multi (Redaction)                — Apache-2.0
  - knowledgator/gliner-linker-large-v1.0 (Disambig) — Apache-2.0
  - knowledgator/gliner-relex-multi-v1.0 (Rel)   — Apache-2.0
  - urchade/gliner_multi-v2.1 (Entity)           — Apache-2.0
  - urchade/gliner_multi_pii-v1 (Redaction)      — Apache-2.0
  - ibm-granite/granite-embedding-311m-multilingual-r2 (Embed, GPU tier, default) — Apache-2.0
  - ibm-granite/granite-embedding-97m-multilingual-r2 (Embed, CPU tier)           — Apache-2.0
  - intfloat/multilingual-e5-small (Embed, fallback) — MIT
  - sentence-transformers/all-MiniLM-L6-v2       — Apache-2.0
  - microsoft/llmlingua-2                        — MIT
  - answerdotai/ModernBERT                       — Apache-2.0
  - google/mt5                                   — Apache-2.0
  - Qwen/Qwen3-* (dense + MoE, all sizes)        — Apache-2.0

OPT-IN, NON-DEFAULT model with license flow-down obligations:

  - google/embeddinggemma-300m (and unsloth mirror) — Gemma license.
    This is NOT a default. The `gemma` license is not OSI open-source: it
    permits commercial use but carries Google's Prohibited Use Policy and a
    flow-down obligation (every downstream recipient must be bound to the Gemma
    use restrictions, with a Notice). The default Embedder is IBM Granite
    Embedding r2 (Apache-2.0; multilingual-e5-small, MIT, remains a fallback). POLICY: the Gemma family is non-default and no Gemma-backed Provider ships
    in this repo; were one offered, it would be an explicitly opt-in Provider carrying
    the Gemma license and its flow-down obligations (not a default, not auto-installed).

Standing rule: a GLiNER/HF checkpoint's license is per-weight and cannot be
inferred from the library. Every new default Provider requires per-model
weights-and-code license verification recorded before it ships (ADR-0012).

================================================================================
Third-party Python runtime dependencies
================================================================================

Every runtime dependency on a DEFAULT install path is permissively licensed.
Two named exceptions sit off that path and are documented in full in
THIRD-PARTY-LICENSES.md: `easydict` (LGPL-3.0, pulled only by the opt-in
`latence-linkpred-ultra[ultra]` extra) and the NVIDIA CUDA wheels pinned by the
two GPU environments under `deploy/env/`, which are proprietary. Full list with
verified licenses in THIRD-PARTY-LICENSES.md. Notable dependencies that
distribute their own NOTICE (Apache-2.0) and whose attribution is preserved here:

  - Apache Arrow / PyArrow  — Copyright The Apache Software Foundation, Apache-2.0
  - Apache Airflow          — Copyright The Apache Software Foundation, Apache-2.0
    (optional extra of latence-runner-airflow only)