Making the invisible decisions of computational science visible, verifiable, and reproducible
Before magnetic resonance imaging, a physician who needed to see inside soft tissue had two options: cut, or guess.
The MRI changed medicine not by making the body less complex, but by making its interior legible without opening it. It turned something opaque into something you could read, measure, and compare.
Probabilistic AI has put science in the pre-MRI position. A model reads data and returns a candidate, a measurement, an interpretation, or a ranking. The output is fluent and often excellent. But the decision behind it is soft tissue: dense, internal, and invisible. Why this result and not the ninety thousand others the model scored? What did it actually see? How confident should we be, on a basis someone other than the model can check? If we re-ran it tomorrow, would we get the same answer, and could we prove it?
Today those questions get answered by cutting, meaning expensive wet-lab or re-run validation on a tiny fraction of outputs, or by guessing, meaning trusting the model's own say-so. Neither scales. Neither satisfies a regulator, a partner, or a reproducibility standard.
Rosicodex is an MRI for AI processes. It does not open the black box, and it does not slow the science. It produces, at the moment of inference, a readable and verifiable image of the decision: what the model saw, how reliable the result is on an independently recomputable basis, and a tamper-evident record a third party can reproduce without re-running the model.
There is a category difference between monitoring AI and instrumenting it. A dashboard sits beside the workflow and reports on it. An instrument sits inside the workflow and produces a primary measurement that becomes part of the result. A chromatogram is not a report about a separation. It is the separation's record. An MRI is not a note about the tissue. It is the tissue, made legible.
Rosicodex produces that kind of record for each AI-assisted decision. It is generated by the apparatus during inference, not narrated afterward, which is precisely why it can be trusted. A record assembled after the fact can be edited. A record produced at the moment of observation and cryptographically bound cannot.
Three things make the image legible:
The model stays a black box. The decision becomes visible.
These are not failures of any particular system. They are structural properties of putting a large probabilistic model inside a workflow that has to answer to someone later. They apply wherever that is true.
A scoring or ranking model filters a very large candidate set before anything reaches a bench, a reviewer, or a report. On what basis did it rank item 1,247 above item 94,832? A receipt captures the model version, the scoring context, and the confidence distribution across the shortlist, so the filter becomes auditable rather than assumed.
Only a small fraction of outputs get confirmed by an orthogonal high-confidence method. For everything else, the this-is-real signal comes from the model itself, or from assays and instruments carrying their own noise. A receipt tags each result by evidence tier, so the basis for belief is visible at a glance rather than flattened into a single number.
Systems that learn from their own results retrain continuously. A result now being acted on was produced by a model that has since changed many times. A receipt pins each result to the exact model version, data snapshot, sampling parameters, and seed that produced it, a frozen record that survives every later update.
Real systems optimize several competing properties at once. Those are trade-offs, and the model's internal weighting of them is not externally visible. A receipt records the objective weighting and constraints in force at decision time, so a partner or reviewer can verify the trade-off matched their context rather than the developer's defaults.
Regulators have moved from principles to specifics. On the therapeutic side, the FDA's risk-based credibility framework for AI used to support regulatory decisions, issued in January 2025 and reinforced by joint FDA and EMA good-practice principles in January 2026, asks sponsors to document context of use, model provenance, versioned code and parameters, independent validation, and audit trails sufficient for an unfamiliar reviewer to judge the model. On the device side, expectations around predetermined change control and post-market monitoring for AI-enabled software functions point the same direction. An instrument running continuously captures that trail prospectively, instead of reconstructing it under deadline.
When an AI-derived result leaves the organization that produced it, whether to a pharma partner, a clinical customer, or a downstream lab, the result transfers but the decision behind it usually does not. A receipt turns the deliverable from a result into a result plus an auditable case: what the model saw, how it scored, what it rejected, and what data grounded it. That is a more defensible commercial asset and a more useful scientific one.
There is a further, and friendlier, point. Where results are contested, the dispute is frequently unresolvable from outside, because the decision itself was never visible to begin with. A provenance record changes the terms, converting a contest of claims into a matter of record. For an organization confident in its science, that is not a threat. It is the ability to prove what you already believe is true.
The same instrument, in three common configurations.
Generative models propose candidates, an experimental system validates a fraction of them, and the results retrain the model. The binding question is which candidates advanced and why, asked months or years later by a regulator or a partner. The receipt is the frozen decision record that outlives the model version that made it.
ML converts raw instrument data into a quantity or identification that customers then act on. Here the receipt functions like an uncertainty budget: it travels with the measurement, states what was actually observed versus inferred, and lets the recipient recompute the reliability rather than accept it on the vendor's authority.
A model produces a determination about a specific subject, sometimes in a setting with no specialist available to overrule it. The record has to hold up years later both to a regulator reviewing the device and to a clinician reconstructing a specific decision. Per-inference provenance is least optional here, and hardest to retrofit.
The model keeps its secrets. The science gets its evidence.
Three forces converge. Probabilistic AI has become a routine instrument inside the scientific workflow rather than a novelty at its edge. Reproducibility is both a recognized crisis and, under recent United States policy on gold-standard science, an explicit expectation for funded and regulated work. And regulators have begun to ask, concretely, for the exact records an inference-time provenance instrument produces.
The gap between how fast AI now makes scientific decisions and how little of that decision is visible has become the field's binding constraint. The organizations that make their AI legible will clear reviews, close partnerships, and defend results that the rest cannot.
Rosicodex wraps any probabilistic AI call as a sidecar, so it competes with neither the model nor the lab system. It makes both auditable. For each invocation it emits a receipt inline: it hashes the exact prompt before the call, pins the exact input records during retrieval, computes a deterministic reliability score from observable properties of the retrieval, captures the output and its hash, and links the receipt into an append-only, hash-chained ledger with a verification routine. Abstentions and blocked inputs are recorded on the same ledger.
The approach is patent pending. A working, model-agnostic system exists today and runs a live demonstration.
An MRI did not make the body simpler. It made the body seen, and in doing so it changed what medicine could safely do. Computational science is at the same threshold. The instrument that makes AI decisions seen, without opening the box and without slowing the work, is the one that lets this field move at the speed of its ambition and still stand behind its results.
Rosicodex is that instrument.
Bring one AI-assisted result your team would least like to have to defend. In twenty minutes we will look at what it would take to make it stand on its own record.