Certanum  /  Pharmaceuticals

Every derived value should have a derivation.

Data-integrity controls govern the record. When an AI system produces a derived value inside that record, a separate question remains: was the figure derived from the declared source data, by the declared computation?

Certanum is being built to answer that question with evidence, for every consequential figure.

01The problem

When a person entered the number, attribution implied derivation. When a generative component produces it, the two come apart.

A record can be attributable, contemporaneous, complete and retained while the derivation of a figure inside it remains unevidenced. An AI system can select the wrong source value, apply the wrong operation, or state a figure it never computed — and the output text looks identical in each case.

Output review cannot separate a computed figure from a generated one. A tool-call log establishes that a calculation ran; it does not establish where the values entering it came from.

Two panels. The upper panel lists the nine ALCOA+ data-integrity attributes and what each establishes about the record. An arrow leads to a lower panel headed Derivation, asking whether the released figure was produced from the declared source data by the declared computation — a question the data-integrity attributes do not answer.
Data-integrity controls govern the record. The derivation of the figure inside it is a separate question.
02Where it applies

Where consequential figures arise.

A figure is consequential when it is acted upon, recorded, submitted or defended.

Clinical development
Derived endpoints, safety figures and statistical outputs that reach a regulatory submission.
Manufacturing & quality
Batch-release, yield, potency and other GxP calculations.
Medical imaging
Measurements an AI model extracts from scans and images, bound to the model version and input by digest.
Pharmacovigilance
Case counts, rates and signal figures derived from safety data.
Precision medicine
Dose and risk calculations for individual patients.
Research & development
Assay, analysis and modelling figures as AI moves into laboratory workflows.
03Existing requirements

The record already has rules. AI adds a question they were not written for.

21 CFR Part 11  ·  FDA
Requirements for electronic records and electronic signatures, including audit trails. They govern the integrity of the record.
ALCOA+  ·  FDA and EU data-integrity guidance
Data should be attributable, legible, contemporaneous, original, accurate and complete. These attributes describe the record; the derivation of a figure inside it is a separate question.
EU GMP Annex 11
Requirements for computerised systems used in GMP-regulated activities.
Draft EU GMP Annex 22  ·  consultation July 2025 ↗
Covers static, deterministic AI models in GMP, and states that generative AI and large language models should not be used in critical GMP applications. Still a draft as at September 2026.
FDA draft guidance  ·  January 2025 ↗
A risk-based credibility assessment framework for AI models used to support regulatory decision-making for drugs and biologics, for a defined context of use. Draft, not for implementation.

Descriptions of regulatory and policy instruments are indicative, are not legal advice, and should be confirmed with qualified counsel. Certanum confers compliance with nothing. It produces evidence an existing assurance, validation or audit process may require.

04What Certanum establishes

For every consequential figure: computed, traceable, reproducible — or withheld.

Computed. The figure is produced by declared, deterministic software — not generated by the model.

Traceable. Every operand entering the computation resolves to an authoritative source value, or to a declared, immutable artifact bound by digest.

Reproducible. A third party can re-execute the declared computation over the recorded operands and obtain the released figure.

Withheld when unsupported. An operand that cannot be resolved does not degrade the answer. It withholds it.

The architecture partitions the generative component out of the numerical path: the model interprets; deterministic software computes. That is the same boundary draft Annex 22 draws for critical GMP applications. Certanum is additional to data-integrity and computer-system-validation controls, not a substitute for them.

05Questions
Does Certanum replace computer system validation or data-integrity controls?
No. Those controls govern the system and the record. Certanum addresses a separate question: whether a consequential figure produced with AI was derived from the declared source data by the declared computation. It is additional to them.
Can a measurement produced by an AI model be evidenced?
Yes, as declared evidence rather than an authoritative source. The model output is bound by digest to the model version and input, so what entered the computation can be established and re-examined.
Does Certanum make a system GxP compliant?
No. Certanum confers compliance with nothing. It produces evidence that an existing validation, quality or audit process may require.
Founding pilots
Two founding pilots — one may be in pharmaceuticals.

One workflow. One class of consequential figure. Success criteria agreed before we begin. Certanum funds the initial pilot work.

Founding pilots →
Other industries  ·  Financial services  ·  Defense  ·  AP-1 standard ↗