Are We There Yet? (Trailer)

Season one read draft EU GMP Annex 22 as the one text that drew a hard line on model type: static and deterministic machine learning only in critical GMP applications, with adaptive models, probabilistic models and generative AI told they should not be used there. On 2 October 2026 EMA published the report of its 30 June expert workshop, and its introduction says the drafting group has discussed opening that scope, provided the models comply with the annex and sit under a fully documented, risk-based control strategy. Sam and Sarah read the fifteen pages closely: industry's case against technology labels, the ICH Q9 vocabulary of uncertainty, importance, complexity and detectability, where the limits of mitigation lie, the report's distinction between human oversight and human-in-the-loop, layered controls that prevent, detect and contain, the lifecycle and supplier questions, and what the drafting group says it may now write. Then the four recurring systems go through the proposed test instead of the old gate, the agent longest of all, and the episode closes on what a site should build now because it is the same evidence package either way.

A site head wants one slide by Friday: four AI systems, three columns, do now, wait for, do anyway. Sam and Sarah close the season by reading the two documents that say where the rulebook is heading: the joint EMA–FDA Guiding Principles of Good AI Practice (January 2026), ten principles on two pages that put context of use, model risk, fit-for-use data, lifecycle and the human–AI team under both logos for the first time; and ICH's October 2025 reflection paper on advanced manufacturing, which names process modelling and continuous process verification as topics, lists the gaps in the 2011 Points to Consider in its own words, and proposes a process-models guideline as the first step, with no topic adopted yet. Then the calendar as the landscape verified it (CDER's promised manufacturing guidance, the Annex 22/11/Chapter 4 finals, the FDA draft, the AI Act's deferred dates), the EU machinery behind it (EMANS 2028, the NDSG and QIG workplans), the sandboxes that exist and the GMP sandbox that doesn't, the six open questions as the season found them, a scorecard for the four recurring systems, and Sam's verdict: nearly there on vocabulary, not on instruments, and the plant floor decides.

A vision vendor pushes a new model version at two in the morning and a language-model provider deprecates the version a site validated. Sam and Sarah split 'change' into four kinds (the environment drifts, the owner retrains, the model learns, the vendor pushes) and read the instruments inside out, anchored on ICH Q12: Q10 and the Annex 11 and Annex 22 drafts inside the quality system; Q12's reporting categories, Established Conditions and the post-approval change management protocol in the filing; FDA's 2025 draft tying them together with a lifecycle maintenance plan; and the device-side predetermined change control plan as the template nobody has married to the PACMP. Then the two questions FDA asked in 2023 and ICH repeated in 2025, a tier-by-tier rule for which instrument applies, the vendor update no document covers, and the four recurring systems taken through all four kinds of change, ending with what an AI PACMP for the soft sensor would actually say.

Every document that lets AI into GMP lets it in on one condition: a human in the loop. Sam and Sarah read what that phrase actually means across six documents, anchored on EMA's 2024 reflection paper: BioPhorum's three oversight settings and the grade each one earns, the EU AI Act's Article 14 list of what an overseer must be able to do, draft Annex 22's operator monitored like any other manual process, FDA's and the GMLP principles' human-AI team, and the paradox the documents state themselves, that oversight is both a control and a risk. Then Sam's question, parked since episode four: which human, with which competence, evaluating which inputs, measured how. The four recurring systems get the three questions, the agentic assistant gets the night-shift test, and Sam's position is offered for the listener to test: encode every check a computer can run, name the human judgment by competence, and if that competence is not on shift, the control is fiction and the grade must not get credit for it.

Every framework this season assumes the data under the model are trustworthy. Sam and Sarah read what the rulebook actually requires: ALCOA and its pluses from FDA's 2018 Data Integrity Q&A, PIC/S PI 041 and the Annex 11 revision; FDA's 2025 'fit for use', which adds relevance to reliability; and the data-governance clauses of the AI Act, EMA's reflection paper and draft Annex 22. The Annex 11 revision turns the training-data pipeline into a GMP computerised system, with validated transfers, unique accounts, audit trails and backup. Then the gap: nobody says how to get from a historian, a batch record and a LIMS to a fit-for-use dataset. The soft sensor's training set is interrogated attribute by attribute, the batch model's golden batches become exclusion decisions, the camera's labels and vials get their clauses, and the agentic assistant's data turn out to be documents, versions and all.