Nouli Policy Pair: Policy extraction and evidence checks

How they work together

Policy flows through the extractor; information flows through the evidence model; Nouli applies the rules and returns TRUE, FALSE or UNKNOWN.

The extractor turns the policy into executable rules. The evidence model checks which required facts the supplied information supports. Nouli applies the rules, including calculations and comparisons, to reach the decision.

Model at a glance

Property	Value
Input	Policy text, case narrative and a query
Output	TRUE, FALSE or UNKNOWN
Components	Policy extractor and evidence model
Tested full-pair runtime	DGX Spark / Linux ARM64 / CUDA
Inference	Local; no hosted API key required
License	Apache-2.0; upstream notices apply

Example input and output

Input

Policy: The taxpayer qualifies if the taxpayer is at least 65 years old on December 31 of the tax year.

Story: Mira was born on April 14, 1950.

Query: taxpayer Mira, other person Tomas, tax year 2025.

Recorded output (using the bundled policy program; trace fields omitted):

Recorded output

{
  "status": "completed",
  "decision": "true"
}

Mira is 75 at the end of 2025, so she meets the policy's age requirement. Example story · Query · Output excerpt.

Quickstart with Nouli CLI

Requires Python 3.10 or newer, the HF CLI, and the documented CUDA environment. The supplied bootstrap pins Nouli CLI 0.5.1. Public base weights are downloaded separately during setup.

Download the complete repository:

Terminal

hf download nouli-ai/nouli-policy-pair --revision 7de8e7f56085068f9cc380be829455b197300ebe --local-dir ./nouli-policy-pair
cd nouli-policy-pair

Complete the runtime setup: build the native tools, open the pinned CUDA container, install dependencies, fetch the bases and generate inference/local/config.json.

Then run the bundled example from the prepared runtime:

Terminal

.live-venv/bin/python inference/live_pair.py \
  --config inference/local/config.json --device cuda \
  --policy inference/examples/policy.txt \
  --story inference/examples/age-eligibility-story.txt \
  --taxpayer 'Mira' --other 'Tomas' --tax-year 2025 \
  --output outputs/live.json

Read outputs/live.json for the decision. To evaluate your own case, replace the policy, story and query values. The setup guide also covers B-only execution with a saved program and offline reuse.

Base models

This repository includes the selected extractor adapter and evidence-model delta with their custom runtime. Keep the complete download together and follow the supplied loader instructions.

Limitations

Evaluated on bounded synthetic policies and narratives. Real-world tax eligibility is unvalidated.
Policy interpretation and evidence checks can make errors. Missing or unsupported facts can produce UNKNOWN.
Full-pair use requires the documented CUDA setup. A clean installation on unrelated hardware remains unverified.

License and citation

Apache-2.0, with upstream notices retained in NOTICE and inference/licenses/.

Nouli AI. Nouli Policy Pair. Cite the published model version and commit.