DAR: Deontic Reasoning with Agentic Harnesses

Researchers tested if giving AI "tools"—such as a search command or a Python interpreter—helps it follow complex laws and policies. They found that top-tier AI models became significantly better at solving legal problems this way. However, smaller, open-source models actually performed much worse. They also wasted far more computational resources.

The bottleneck of static context

Deontic reasoning is the process of answering questions by applying explicit rules or policies to specific facts. Examples include calculating tax liability under a statute or determining the outcome of an immigration appeal. The fundamental technical challenge lies in the nature of the rulesets themselves. Statutes are often long, dense, and heavily cross-referenced. A single obligation might be modified by a definition located hundreds of lines away.

The standard approach in current LLM evaluation is "direct reasoning." In this setup, the entire statute, the case facts, and the question are bundled into a single prompt. As shown in, the model must ingest everything in one pass.

It is expected to identify the relevant provisions immediately. This puts immense pressure on the model's long-context capabilities (the ability to process large amounts of text at once). If the model fails to locate a specific exception, the entire reasoning chain collapses.

Interactive exploration via DAR

To address this, the authors introduce Deontic Agentic Reasoning (DAR). Instead of treating the statute as a static block of text, DAR treats it as a queryable resource. In this setup, the statute is stored as a file within a sandboxed environment (an isolated computing space). The model is given access to a "harness" of general-purpose tools.

The mechanism works through an iterative loop: 1. The model receives only the case facts and the question. It also receives instructions on how to use the environment. 2. To find the necessary rules, the model issues tool calls. It might use grep to search for keywords or sed to extract specific lines. 3. The output of these tools is appended to the conversation history. This serves as new context for the next step. 4. The model may also execute Python code. This handles complex numerical computations required by the rules.

By interacting with the text incrementally, the model navigates cross-references. It can search for definitions only when a term is encountered.

Capability amplification and token inflation

The impact of this shift depends entirely on the underlying intelligence of the model. The authors evaluate several models using the DeonticBench benchmark. This includes difficult tasks involving U.S. tax law (SARA), immigration (USCIS), and airline policies.

The paper reports that frontier models benefit significantly from this agentic approach. For example, under the Terminus-KIRA harness, GPT-5.2's accuracy on SARA-Numeric jumps from 30.3% to 60.0%. This means the model doubled its success rate on math-heavy tax tasks. Similarly, Claude Sonnet 4.5 increases its SARA-Numeric accuracy from 36% to 54%. These results, summarized in, suggest that the harness unlocks latent reasoning abilities.

These abilities were previously obscured by the "noise" of a massive, static prompt.

However, the results for open-source models tell a different story. The authors observe a sharp decline in performance for weaker models. On the SARA-Numeric task, Qwen3.5-35B drops from 34% accuracy to just 11% when using the Terminus-KIRA harness. Even more striking is the "collapse" seen in the Airline task. Every open-source model tested dropped to near-zero accuracy.

This degradation comes at a heavy computational cost. As illustrated in, open-source models consume up to 4× more tokens per trial than frontier models.

Tokens are the basic units of text processed by an LLM. The authors suggest that the harness acts as a "confidence amplifier" for these models. Instead of finding the right rule, the extra turns allow the model to build longer, more confident, but incorrect answers.

Limits of the agentic scaffold

While the DAR framework shows promise, the study highlights several critical limitations. First, the current implementation relies on manual terminal commands like grep and sed. The authors note that very large regulatory corpora pose a problem. For example, the full U.S. Internal Revenue Code is massive. Even frontier models would struggle to "read" enough text to find a single provision. A more scalable solution would require a hierarchical retrieval system. This would narrow the search space before the agent begins reasoning.

Second, the scope of the evaluation is limited to three specific domains. These are U.S. tax, immigration, and airline policies. Legal structures vary wildly across different jurisdictions. It is unclear if these agentic gains would generalize to criminal law or international treaties. Finally, the study uses general-purpose terminal tools. The authors suggest that specialized "provision-aware" tools might yield different results. Such tools would be designed specifically to navigate legal hierarchies.

Verdict: A tool, not a cure

Is agentic reasoning the future of legal AI? The answer depends on your model.

If you work with frontier-class models, DAR is a clear win. It turns a passive reading task into an active investigation. However, using agentic workflows with smaller, open-source models carries risks. The authors' findings serve as a warning. In those cases, the harness does not solve the reasoning problem. It merely subsidizes the error with higher token costs and misplaced confidence.

For practitioners, the takeaway is clear. Do not assume that adding tools will compensate for a lack of reasoning capability. Until open-source models match the "judgment" required to use these tools, agentic scaffolds remain a luxury. Code for the DAR setup is reportedly available; see the paper for the canonical link.