Metadata-Version: 2.4
Name: arc-court
Version: 0.1.0
Summary: Adversarial Research Court — tiered, cited claim verification built on recurve.
Keywords: llm,fact-checking,claim-verification,agents,citations,recap
Author: ned
License-Expression: MIT
License-File: LICENSE
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: Operating System :: OS Independent
Classifier: Programming Language :: Python :: 3 :: Only
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Typing :: Typed
Requires-Dist: recurve>=0.1.0
Requires-Dist: tavily-python>=0.7
Requires-Dist: typer>=0.26
Requires-Dist: rich>=14
Requires-Dist: datasets>=4.0 ; extra == 'bench'
Requires-Dist: deepeval>=4.0.5 ; extra == 'eval'
Requires-Dist: ragas>=0.4.3 ; extra == 'eval'
Requires-Dist: bespokelabs ; extra == 'hosted-nli'
Requires-Dist: recurve[openai] ; extra == 'local'
Requires-Dist: torch>=2.2 ; extra == 'nli'
Requires-Dist: transformers>=5.0 ; extra == 'nli'
Requires-Dist: sentencepiece>=0.2 ; extra == 'nli'
Requires-Dist: chromadb>=1.0 ; extra == 'search'
Requires-Python: >=3.11
Project-URL: Homepage, https://github.com/main-path/aethel
Project-URL: Source, https://github.com/main-path/aethel
Project-URL: Issues, https://github.com/main-path/aethel/issues
Provides-Extra: bench
Provides-Extra: eval
Provides-Extra: hosted-nli
Provides-Extra: local
Provides-Extra: nli
Provides-Extra: search
Description-Content-Type: text/markdown

# arc-court

> The Adversarial Research Court — tiered, cited claim verification.

`arc-court` (imported as `arc`) verifies claims and emits a report with per-claim
**adversarial-survival** scores. It is built entirely on [`recurve`](https://pypi.org/project/recurve/)'s
recursive planning primitives.

**Tiered by design** — because the research is clear that naive multi-agent debate frequently fails to
beat single-agent baselines at equal compute:

- **Tier 0 — claim graph:** extract only *verifiable* claims, decompose + decontextualize into atomic units.
- **Tier 1 — Falsification Gate (default):** retrieve evidence (Tavily / domain sources), a single
  Prosecutor proposes cited falsifiers, a cheap NLI judge scores entailment, numeric/temporal claims
  escalate to a stronger judge. `NOT_ENOUGH_INFO` is a first-class verdict.
- **Tier 2 — Court (opt-in):** the full Prosecutor / Defender / Jury trial, reserved for high-stakes,
  mid-confidence claims — and gated by an **anti-theater** eval that it must beat single-agent
  CoT + self-consistency at equal compute, or it stays disabled.

Part of the [AETHEL](https://github.com/main-path/aethel) project. MIT licensed. Status: alpha.
