Try the audit layer in your browser.
Three tools. The workbench, the readiness check and the keyword search run entirely on this page: nothing is uploaded. The AI draft in the workbench is simulated.
What you will see: fix the scope, the data and the environment, capture an AI draft, record a human review, then replay the fingerprints. If anything changed after the review, the review is marked stale and has to be redone. Every step is written to the evidence log on the right.
1. Scope the task
Record what the assistant may do and how the result will be used.
Illustrative rule: one named reviewer is enough for this context of use.
2. Fix the data and environment
Fingerprints tie the run to an exact dataset and software environment, before any code is drafted or reviewed.
3. Capture the AI draft
The prompt, the model and the output are logged as one entry, with a separate fingerprint for the prompt and for the output.
Prompt (fixed in this demo)
Draft
The draft appears here once generated.
Simulated. The code is canned and no model is called. One of the three drafts contains a planted error: read the code before you decide.
4. Record the review
A named person accepts, edits or rejects the draft. Edits and rejections need a reason. The decision joins the chain. Reviewers must differ from each other and from the person who scoped the task. Names are not authenticated in this demo.
5. Replay check
Recompute the fingerprints of the code, the data and the environment and compare them with the logged ones. This demo does not run R. A real replay re-runs the code and compares the outputs as well.
Run it once with both boxes cleared to see a clean pass, then with a box ticked to see a difference.
Press Tamper on any entry to change it after the fact and watch verification fail.
Hash chains reveal edits; they do not prevent them. A real deployment also keeps the latest hash outside the system. Times are UTC (the Z): Brussels is one or two hours ahead. This is a browser demo, not regulog output.
Where does your AI-assisted workflow stand?
One question per principle in the January 2026 FDA and EMA guiding principles of good AI practice. Answer for the workflow you have today. Answer Partly if the practice exists for some of your work or only part of the question is true.
Self-assessed and indicative. A conversation starter, not an audit.
Score: Yes = 2, Partly = 1, No = 0, summed over the ten questions and shown as a percentage. The labels switch at 50% and 80%; these cut-offs are conventions of this demo, not thresholds from any regulator.
Ask the evidence map
Searches a fixed set of answers written for this site. The search is keyword retrieval, not a chatbot, and it does not call a model.
Want these on your own workflow?
Describe where AI is entering your statistical programming and what your QA group asks for.