Attribution

Studycase runs inference against a mix of third-party models. This page lists every one the code actually references, its license, and — for the CC-BY-4.0 models specifically — the mandatory attribution Creative Commons requires (model name, source, license link, a note on what we changed).

Method (no fabrication): every license below was pulled live from the model's Hugging Face repo (GET https://huggingface.co/api/models/<id>, cardData.license / tags) or, for the two NVIDIA Open Model License models and the Qwen license, the repo's own LICENSE file — not recalled from training memory. Access date for every verdict on this page: 2026-07-05. Where a claim could not be pinned to a live source, it is marked UNVERIFIED rather than guessed. The CC-BY verdicts for parakeet-tdt-0.6b-v3 and diar_streaming_sortformer_4spk-v2 are carried forward from the deeper audit in docs/LEGAL-1-NVIDIA-LICENSE.md, which quotes the operative license text directly — this page re-derives the same license tags independently via the HF API as a cross-check and adds every non-NVIDIA model the code references (embedding models, LLM tiers) that LEGAL-1 was not scoped to cover.

Scope note: there is no automated dependency-license manifest yet (that's ticket LEGAL-9, still open). This page is a manual inventory built by grepping services/ for every model id actually wired into code (services/asr, services/rag, services/inference) plus the LEGAL-1 NVIDIA audit. If LEGAL-9 ships an automated scan, reconcile this page against it rather than assuming this list is exhaustive forever.

1. CC-BY-4.0 models — mandatory attribution

These are used under Creative Commons Attribution 4.0, which conditions use on giving credit: model name, creator, license link, and a note on whether the model was modified.

Speech-to-text (fallback / mlx tiers) — NVIDIA, parakeet-tdt-0.6b-v3 and parakeet-tdt-0.6b-v2, licensed under CC BY 4.0. Used unmodified via NVIDIA NeMo (buffered/chunked inference wrapper only — no weight changes); the v3 checkpoint is also consumed via its Apple-Silicon repackaging, mlx-community/parakeet-tdt-0.6b-v3, run unmodified through the third-party parakeet-mlx library (Apache-2.0, not itself a model).

Speaker diarization — NVIDIA, diar_streaming_sortformer_4spk-v2, licensed under CC BY 4.0. Used unmodified via NVIDIA NeMo.

All four verified live via the HF models API on 2026-07-05: cardData.license: "cc-by-4.0", tag license:cc-by-4.0, for every id above.

2. NVIDIA Open Model License Agreement (OML) — not CC-BY, goodwill credit only

Not legally required for our deployment (we run inference server-side and never redistribute the weights, so the OML's Notice-file/redistribution duty doesn't attach — see LEGAL-1 §6 for the full reasoning), but credited here anyway for completeness and goodwill.

ASR (primary streaming tier) — NVIDIA, nemotron-speech-streaming-en-0.6b (or parakeet-unified-en-0.6b), used under the NVIDIA Open Model License Agreement. Used unmodified.

3. Permissively-licensed models used in the RAG / LLM pipeline

Not CC-BY, no attribution legally required (MIT/Apache-2.0 have no attribution-on-the-product clause the way CC-BY does), but listed here for completeness since the code depends on them directly.

4. Other models referenced (non-permissive / custom license)

5. Explicitly out of scope for this page

• Anthropic Claude (Sonnet/Haiku tiers) — accessed via the Anthropic API as a hosted service, not a downloadable model with a license file; no attribution obligation of the CC-BY kind applies. Not a "model asset," so excluded from the table above.
• `ocr_via_vlm` / Qwen-VL OCR escalation (services/inference/studycase_gateway/casefiles.py) — this is a documented stub (raise NotImplementedError) describing a future pattern; no VLM model is actually downloaded or called by this code today. Nothing to attribute yet — re-visit this page when ASR/OCR-VLM work lands a real backend.
• Supporting libraries that are not model weights (NVIDIA NeMo toolkit, Apache-2.0; parakeet-mlx, Apache-2.0; eyecite/courts-db, BSD/MIT family) are software, not CC-BY assets, so they aren't repeated here as a full scan — that's LEGAL-9's job (dependency license scan across all pip packages), still open.

No entries on this page are UNVERIFIED as of the 2026-07-05 audit — every license tag was independently confirmed against the HF models API (or the repo's own LICENSE file for the two non-HF-standard licenses) at that date. If a future model is added to the code without updating this page, treat its license as unverified until someone does the same check.

6. App-footer wiring — NOT done here, tracked separately

This document is the content LEGAL-2 asked for. The ticket also asks for it to be "linked from app footer and docs." The docs half is satisfied by this file living in docs/ and being linked from docs/LEGAL-1-NVIDIA-LICENSE.md and the ticket board. The app-footer link is `apps/web` work, owned by a different lane, and remains outstanding. Exact link target for whoever picks that up:

• Link text: e.g. "Model attributions" / "Third-party licenses"
• Link target: /docs/ATTRIBUTION.md if apps/web serves the docs/ tree directly, otherwise render this file's content at a stable app route (e.g. /attributions) and keep it in sync with this source file — do not fork the content.
• Related open tickets that should point here once built: DOCS-11 (FAQ "model attribution links"), DEMO-11 (landing page "model attributions in footer").