AI Reproducibility

aicr is a reproducibility harness for LLM deployment testing. It drives multiple model deployments in parallel, runs a set of prompts a configurable number of times across each, and records cryptographically signed, hashed evidence for every response.

Built to support the Clinical-AI Reproducibility Annex.

What it does

  • Authenticates with Microsoft Entra ID (DefaultAzureCredential) for Azure deployments, or an API key for OpenAI-compatible APIs.
  • Sends fully parameterised request bodies assembled from config/deployments.json.
  • Captures the full raw JSON response for every run to its own file under output/<timestamp>/runs/.
  • Computes a SHA-256 of the assistant message content as the identity key — produces identity groups automatically across runs.
  • Scores list tasks (exact matches, coverage, hallucinations, ordering accuracy) and title-recall tasks.
  • Writes manifest.json, manifest.csv, summary_counts.csv, summary_pct.csv.

Get started

Download the desktop app or CLI binary and follow the Your first project guide. Experiments can be configured, run, and analysed either from the command line or in the desktop GUI.

Browse results

The Results viewer is hosted at ai-copyright-reproducibility.tauheed-elahee.com/viewer/.

Sections

Section What’s inside
Getting Started Download binaries, build from source, run your first experiment
Project layout Directory structure, configuration files, and deployment modes
Input text.json, queries.json, prompts.json — what to ask and who to ask it to
Running CLI reference, helper scripts, and troubleshooting
Desktop GUI The aicr-gui desktop app — configure, run, and analyse experiments visually
Output Output files, manifest columns, scoring algorithm, and interpreting results
Results viewer Hosted Blazor viewer — load local or default manifests
Methodology Hashing rationale, system_fingerprint drift, identity groups
Glossary Definitions for every key term

This site uses Just the Docs, a documentation theme for Jekyll.