An MRI for AI-assisted science
Your AI's results, still standing after the model changes.
At the moment a model reads your data, Rosicodex records what it saw, how it was configured, and what it decided. Anyone outside your team can verify that result later, without re-running the model and without taking your word for it.
01 / The rerun
Models retrain. The result you reported last year still has to hold.
Move through time and switch the receipt on and off. This is the question a partner, a regulator, or a reviewer eventually asks.
DAY 1 / MODEL V3
The model ranks candidate C-17 first.
The team advances C-17. The result goes into a report, a partner deck, and the program plan.
WITH A ROSICODEX RECEIPT
Recorded at the moment of inference.
- inputs and sources pinned
- model, version, and seed captured
- output hashed and chained
Illustration.
02 / Tamper test
Change one old record and the chain tells on it.
Every receipt carries the fingerprint of the one before it. Try editing receipt 3.
All five receipts verify against the chain.
03 / How it works
An instrument inside the workflow.
Before MRI, a physician who needed to see inside soft tissue could cut or guess. Generative AI has put science in the same position. A model returns a fluent, often excellent interpretation, and what it saw, how sure it should be, and whether tomorrow's rerun would agree all stay out of view.
Rosicodex leaves the model closed. The receipt opens before the model is called and seals before the answer returns, so nothing is reconstructed after the fact.
- GuardsInjection and answerability checked before the model runs.
- RetrievalThe exact records pinned and hashed before the call.
- ConfidenceRecomputable by anyone, never the model's self-report.
- DecisionAnswer, caveat, or route to human review.
- ReceiptSealed into a tamper-evident ledger.
04 / Where it matters
Someone outside your team will ask. Here is who, and when.
WHO ASKS
A partner's diligence team, a licensee, or an IND reviewer.
WHEN IT BITES
A candidate came from a model that has since been retrained, and someone asks why it was chosen.
WHAT THE RECEIPT ANSWERS
Exactly what the model saw and how it was configured when it ranked that candidate.
05 / Deploying Rosicodex
Built to wrap the pipeline you already run.
Rosicodex sits at your inference step and works with any model. Every deployment is fitted to one real workflow and the date it has to be ready by.
- 1Twenty-minute callWhich result has to stand up, to whom, and by when.
- 2Scope one workflowA written plan worked back from your deadline.
- 3IntegrateRosicodex wraps your existing inference call, in your cloud or on your own machines. Your data is handled under your company's rules and compliance standards.
- 4Receipts in productionEvery result ships with its record, sealed into a tamper-evident ledger.
- 5Verification reportA reviewer-ready account of what was recorded and how anyone can check it.
06 / The standard
Gold Standard Science (Executive Order 14303) asks for results that are reproducible, transparent, and honest about uncertainty. FDA's January 2025 draft guidance on AI in drug and biological product submissions asks for model context, data provenance, and an audit trail. The receipt turns those asks into fields.
07 / Status
- Working instrument. A live demo on real model calls.
- Evaluated. A golden-case evaluation suite that runs on every change.
- Patent pending. Rosicodex LLC.
- NSF SBIR. Pending.
08 / Founder
Andy Rosic founded Rosicodex after two decades building products where the record has to be right, including for Apple Retail, Dun & Bradstreet, and Cross Country Mortgage.
Bring the AI result you would least like to have to defend.
Twenty minutes, and we will look at what it would take to make it stand on its own record.