ABDA-NL turns facts, assumptions, and rules into arguments you can inspect, question, and revise in your browser.
Tip
Try it now at demo.abda-nl.org. You can explore the built-in and community scenarios immediately. Registration with an email code takes about 30 seconds. The first 100 registered users can activate $5 of free AI credit, enough for roughly 200 typical LLM calls. The exact number depends on the chosen model and the length of each exchange.
ABDA-NL makes structured argumentation approachable without hiding the formal reasoning underneath. The ABDA engine computes the arguments, attacks, and conclusion labels. The optional AI assistant helps explain those results in natural language, find relevant source passages, and draft changes for human review.
| Follow the reasoning | See which conclusions are accepted, rejected, undecided, or absent, then inspect the supporting and attacking arguments. |
| Ask what changed | Toggle assumptions and defeasible rules, preview the effects, reset the scenario, or undo the reset. |
| Talk with the scenario | Ask grounded questions, open highlighted evidence, stop a slow response, retry an answer, or edit and fork the conversation. |
| Build your own case | Create a scenario in a guided editor, write ASPIC- rules directly, or import YAML and JSON. |
| Bring source material | Attach text, Markdown, and text-based PDFs as context for explanations and citations. |
| Keep work portable | Save private scenarios, share a read-only view, publish a community scenario, or export a self-contained JSON file. |
| Work through an agent | Connect Codex or Claude Code through scoped MCP access to inspect, create, import, and revise scenarios. |
| Present without an LLM | Choose AI off and keep the complete deterministic explorer available. |
The interface is designed for argumentation researchers, students, teachers, and anyone curious about how reasons support or challenge a conclusion. You do not need to know ASPIC- notation to begin.
- Open the live demo and choose a scenario from the title menu.
- Select Explain beside a conclusion to inspect its derivation.
- Change an assumption or rule and preview how the conclusion labels respond.
- Ask the AI assistant why a conclusion holds, or request supporting sources.
- Open New scenario or Import scenario file when you are ready to try your own example.
AI is optional throughout this flow. It can describe the formal result and suggest edits, but it never decides which conclusions follow. Proposed edits must pass validation and remain under the user's control.
Gemini 3.8 Flash is the default funded model because it offers a strong balance of capability, speed, and cost. Other qualified Gemini, Claude, and GPT models are available from the same menu.
For registered users, CloudBank supplies the primary funded route. Eligible provider outages can fall back to the same model through OpenRouter. You can also bring your own Anthropic, OpenAI, Google, or OpenRouter key. A personal key stays in the current browser tab and is not stored by the service.
Choose AI off from the model menu, or from Account > AI access, for a manual presentation. Scenario exploration, formal explanations, editing, import, and export continue to work without a model call.
The guided editor lets you describe statements and connect them with strict or defeasible rules. The Rule text view exposes the corresponding ASPIC- syntax. Both views edit the same knowledge base.
You can add reference documents, preview computed results before saving, and export the complete scenario with its glossary and source text. Imports accept complete ABDA-NL YAML or JSON files, as well as separate rules, glossary files, and supporting documents.
See Your own scenarios for a short walkthrough and the supported formats.
ABDA-NL supports Python 3.10 through 3.13.
git clone https://github.com/idaks/ABDA-NL.git
cd ABDA-NL
python3 -m venv .venv
make install
.venv/bin/abda-nlThe last command starts the app on your computer and opens a browser. When no LLM configuration is present, it starts as a complete deterministic demo. You can also choose the mode explicitly:
.venv/bin/abda-nl --basic # deterministic explorer, no LLM
.venv/bin/abda-nl --llm # require configured LLM access
.venv/bin/abda-nl --no-browser # start without opening a browserFor development and tests:
make install-dev
make testOn NCSA Delta, run demo. The shared launcher manages the process and the
loopback tunnel described in this repository's AGENTS.md.
- Scenario authoring, imports, and exports
- Codex and Claude Code through MCP
- COMMA 2026 demonstration playbook
- Model selection and routing
- Architecture and design decisions
- Service operation and deployment
- PostgreSQL backup and recovery
If you consider ABDA-NL useful for research, teaching, or exploration, please star ⭐ this repository. It helps other people find the project.
ABDA-NL was created by Shawn Bowers, Martin Caminada, Haoyang Liu, and Bertram
Ludäscher for the COMMA 2026 demonstration track. Citation metadata is provided
in CITATION.cff.
The project combines a natural-language workspace with the
ABDA engine for deterministic
argumentation semantics. See THIRD_PARTY_NOTICES.md
for component attribution and corresponding-source information.
This distribution is licensed under
GNU GPL version 3 (GPL-3.0-only). The original MIT notice is
preserved in LICENSES/ABDA-NL-MIT.txt. The paper
and cited research material are not relicensed by this repository.