This is the fact check of the episode as published. Each factual claim was extracted from the script and verified by an AI agent that saw only the claim and the primary document (the PDFs in the corpus). Verdicts: SUPPORTED (the document says it), PARTIAL (supported with a difference, noted), NOT-IN-CORPUS (the source is not among the primary documents on disk), NOT-CHECKABLE (an estimate or a characterisation). Opinions voiced by the hosts are listed but not verified. This report is itself AI output and can be wrong; corrections: jack@jackprior.ai.
These statements stand, but the primary documents on disk do not themselves confirm them: they come from the landscape reference (a working index whose rows are checked against the web), are estimates, or are hedged in the episode as such.
| Line | Speaker | Claim | Verdict | Why it stands | Evidence |
|---|
| 10 | SARAH | The landscape reference covers about forty documents, has a lineage diagram and a timeline, and was last updated at the end of August 2026. | PARTIAL | Landscape self-description; it was corrected again on 6 Sep 2026, a week after the date spoken | Landscape header: 'Started 27 Aug 2026'; Views file carries the lineage map and timeline |
| 10 | SARAH | The joint EMA-FDA guiding principles were published on 14 January 2026. | NOT-IN-CORPUS | Document is dated 'January 2026'; the day comes from the landscape (web-verified) | Document dated 'January 2026'; day from the landscape row |
| 10 | SARAH | The joint EMA-FDA guiding principles are final and are principles rather than guidance. | PARTIAL | Published, dated document that calls itself guiding principles which 'may help inform' guidelines | p.1: 'may help inform regulatory policies and regulatory guidelines in different jurisdictions' |
| 10 | SARAH | ICH's reflection paper on advanced pharmaceutical manufacturing was published in March 2026. | NOT-IN-CORPUS | Landscape row is web-verified (27 Aug 2026); the detail is not in the primary document on disk (ICH URL path carries 2026-03) | Landscape row: 'published Mar 2026'; ICH URL path 2026-03 |
| 14 | SARAH | In the United States the relevant law layer is Title 21 parts 210 and 211, and Part 11 on electronic records. | NOT-IN-CORPUS | Names the US statutory layer; 21 CFR is not in the corpus | General naming of the US statutory layer; 21 CFR not in corpus |
| 16 | SARAH | Within ICH documents, 'impact' is used with at least two different meanings. | NOT-CHECKABLE | Characterisation of vocabulary, not a checkable fact: 'model impact' in the Points to Consider and M15 versus 'impact' in Q9-style risk assessment | Vocabulary characterisation across ICH texts |
| 21 | SARAH | FDA published a discussion paper on AI in drug manufacturing in March 2023. | PARTIAL | PDF title page says only '2023'; landscape row is web-verified (27 Aug 2026); the detail is not in the primary document on disk | Title page: 'Discussion Paper \| 2023'; month from the landscape row |
| 21 | SARAH | The current Annex 11 on computerised systems has been in force since 2011. | NOT-IN-CORPUS | Current Annex 11 is not in the corpus; date matches the landscape lineage map | Current Annex 11 not in corpus; landscape lineage map gives 2011 |
| 23 | SARAH | The Annex 22 consultation ran from 7 July 2025 to 7 October 2025. | NOT-CHECKABLE | Hedged in-script as 'per the landscape'; the draft text carries no consultation dates | Landscape row: 'consultation 7 Jul – 7 Oct 2025' |
| 25 | SARAH | The ICH Quality Implementation Working Group's Points to Consider for Q8, Q9 and Q10 dates from 2011. | PARTIAL | Reference number EMA/CHMP/ICH/902964/2011; the copy on disk is the EMA transmission edition dated February 2012 | Header: 'February 2012, EMA/CHMP/ICH/902964/2011' |
| 36 | SARAH | The European Commission published draft Annex 22 on artificial intelligence in July 2025. | NOT-CHECKABLE | Draft text carries no publisher or date; landscape row is web-verified (27 Aug 2026); the detail is not in the primary document on disk | Draft text carries no publisher or date; landscape row web-verified |
| 36 | SARAH | Draft Annex 22 was drafted by EMA's GMP and GDP Inspectors Working Group and co-published with PIC/S. | NOT-IN-CORPUS | Hedged in-script as 'per the landscape'; Annex 22 names no drafting body | Landscape row: 'Drafted by EMA GMP/GDP Inspectors Working Group; co-published with PIC/S' |
| 43 | SARAH | The CSA guidance was drafted in 2022. | PARTIAL | The Feb 2026 text gives only the docket number FDA-2022-D-0795; its guidance history starts at Sept 2025 | Preface: docket FDA-2022-D-0795 |
| 45 | SARAH | The Annex 11 revision was released for consultation in July 2025, alongside Annex 22. | NOT-CHECKABLE | Draft carries no date; landscape row is web-verified (27 Aug 2026); the detail is not in the primary document on disk | Draft carries no date; landscape row web-verified |
| 45 | SARAH | The current Annex 11 is short. | NOT-IN-CORPUS | Current Annex 11 is not in the corpus; no number is claimed | Not in corpus; no page count claimed |
| 45 | SARAH | A revised Chapter 4 on documentation was issued in the same July 2025 consultation. | NOT-CHECKABLE | Draft carries no date but cross-references the Annex 11 and Annex 22 revisions; landscape row is web-verified (27 Aug 2026); the detail is not in the primary document on disk | Preamble ties the revision to Annex 11; no date in the text |
| 45 | SARAH | Finals for Annex 11/22 and Chapter 4 are expected late 2026 with implementation early 2027. | NOT-CHECKABLE | Estimates, hedged in-script | Landscape watchlist estimates |
| 60 | SARAH | No new ICH topic on process models has been adopted following the reflection paper. | NOT-CHECKABLE | Consistent with the paper's own framing ('would begin with a new topic proposal') | Consistent with p.9: 'would begin with a new topic proposal' |
| 60 | SARAH | BioPhorum published AI risk guidance for the pharmaceutical industry in June 2026. | PARTIAL | Document footer is dated May 2026; the landscape gives the release as 2 June 2026 | Footer: '©BioPhorum Operations Group Ltd \| May 2026' |
| 65 | SARAH | No document in the corpus says what a vendor update of a foundation model outside change control does to a validated state. | NOT-CHECKABLE | Absence claim; nothing found across the corpus, nearest hook is Annex 11 draft ch. 7 (Supplier and Service Management) | No passage found across the corpus; nearest is Annex 11 draft ch. 7 |
| 72 | SARAH | 21 CFR Parts 210/211 and the EU GMP guide make deviation investigation and batch record review the quality unit's responsibility. | NOT-IN-CORPUS | General statement of 21 CFR 211 / EU GMP Chapter 1 responsibilities; not in the corpus | General statement; 21 CFR 211 and EU GMP Chapter 1 not in corpus |
| 74 | SARAH | ISPE's GAMP Guide on artificial intelligence, from July 2025, is the industry reference the landscape lists for AI across a GxP lifecycle. | NOT-IN-CORPUS | Paid document, not on disk; attributed to the landscape in-script | Landscape row: 'GAMP Guide: Artificial Intelligence, ISPE, Jul 2025' |
| 80 | SARAH | CDER has a guidance on AI/ML quality considerations in pharmaceutical manufacturing on its 2026 guidance agenda, not yet published. | NOT-IN-CORPUS | Hedged in-script as the landscape's watchlist | Landscape watchlist: 'On 2026 CDER agenda' |
| 85 | SARAH | The landscape's watchlist lists six items as spoken. | PARTIAL | The script's own selection of six from the landscape's eight rows | Landscape watchlist has eight rows; the script selects six |
| 85 | SARAH | Per the landscape, an amending regulation in July 2026 pushed the AI Act's high-risk obligations back to December 2027 and August 2028. | NOT-IN-CORPUS | Amending regulation not on disk; attributed to the landscape in-script; the AI Act text on disk still says 'apply from 2 August 2026' | Landscape §3.3: Digital Omnibus on AI, Reg. (EU) 2026/1744 |
| 88 | SARAH | PIC/S's data integrity guide (2021) sets out the ALCOA-plus expectation. | PARTIAL | Document writes 'ALCOA+'; spoken as 'ALCOA-plus' | PI 041-1 7.4: 'ALCOA+' |
| Line | Speaker | Claim | Verdict | Evidence |
|---|
| 10 | SARAH | The joint EMA-FDA document is titled Guiding Principles of Good AI Practice in Drug Development. | SUPPORTED | Title: 'Guiding principles of good AI practice in drug development', January 2026 |
| 10 | SARAH | The joint EMA-FDA guiding principles document is two pages long. | SUPPORTED | Two pages (one page break before 'Principles') |
| 52 | SARAH | The joint principles call themselves initial collaborative work; per the landscape it is the first time the two regulators have put their names to one set of AI principles for medicines. | SUPPORTED | p.1: 'this initial collaborative work can inform our broader international engagements' |
| 52 | SARAH | The joint principles document contains ten principles. | SUPPORTED | p.1: 'These 10 guiding principles' |
| 52 | SARAH | The joint principles' scope is AI used to generate or analyse evidence across the drug product life cycle, listing nonclinical, clinical, post-marketing and manufacturing phases. | SUPPORTED | p.1: 'used to generate or analyse evidence across the drug product life cycle, including nonclinical, clinical, post-marketing, and manufacturing phases' |
| 54 | SARAH | Principle: human-centric by design. | SUPPORTED | Principle 1, p.2 |
| 54 | SARAH | Principle: a risk-based approach with validation and oversight proportionate to 'the context of use and determined model risk'. | SUPPORTED | Principle 2, p.2, verbatim |
| 54 | SARAH | Principle: adherence to standards, including GxP. | SUPPORTED | Principle 3, p.2 |
| 54 | SARAH | Principle: a clear context of use, defined as role and scope. | SUPPORTED | Principle 4, p.2: '(role and scope for why it is being used)' |
| 54 | SARAH | Principle: multidisciplinary expertise across the lifecycle. | SUPPORTED | Principle 5, p.2 |
| 54 | SARAH | Principle: data governance and documentation, with provenance and processing steps traceable in line with GxP. | SUPPORTED | Principle 6, p.2 |
| 54 | SARAH | Principle: model design and development practices using data that are fit for use. | SUPPORTED | Principle 7, p.2: 'data that is fit-for-use' |
| 54 | SARAH | Principle: risk-based performance assessment of 'the complete system including human-AI interactions'. | SUPPORTED | Principle 8, p.2, verbatim |
| 54 | SARAH | Principle: life cycle management under a risk-based quality system, with scheduled monitoring and periodic re-evaluation, and it names data drift. | SUPPORTED | Principle 9, p.2: 'scheduled monitoring and periodic re-evaluation … (e.g., to address data drift)' |
| 54 | SARAH | Principle: clear, essential information for users in plain language. | SUPPORTED | Principle 10, p.2 |
| 56 | SARAH | The joint principles describe themselves as laying a foundation for good practice and identifying areas where regulators and standards bodies could collaborate, which may inform regulatory policies and guidelines. | SUPPORTED | p.1: 'lay the foundation for developing good practice … may help inform regulatory policies' |
| 74 | SARAH | In the joint principles, principle eight is performance assessment of the complete system including human-AI interactions. | SUPPORTED | '8 Risk-based performance assessment … the complete system including human-AI interactions' |
| 88 | SARAH | Principle six of the joint document says data source provenance, processing steps and analytical decisions should be documented in a detailed, traceable and verifiable manner. | SUPPORTED | Principle 6, p.2, verbatim |
| 90 | SARAH | Principle five of the joint document asks for multidisciplinary expertise covering both the AI technology and its context of use. | SUPPORTED | Principle 5, p.2 |
| Line | Speaker | Claim | Verdict | Evidence |
|---|
| 8 | SARAH | Draft Annex 22 contains a gate (restriction) on generative models. | SUPPORTED | Scope, lines 20-21: 'does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications' |
| 14 | SARAH | In Europe the GMP guide with its annexes is EudraLex Volume 4. | SUPPORTED | 'EudraLex' appears in Annex 15 and the ICH reflection paper |
| 16 | SARAH | 'Critical GMP application' is a term used by draft Annex 22. | SUPPORTED | Scope lines 14-15, 21; 8.1 line 126: 'critical GMP applications' |
| 36 | SARAH | Draft Annex 22 is a new GMP annex written specifically about AI. | SUPPORTED | 'Reasons for changes: Not applicable (new annex)' |
| 36 | SARAH | Draft Annex 22's scope is static, deterministic machine learning models in critical GMP applications. | SUPPORTED | Scope lines 6-16: 'critical applications with direct impact on patient safety, product quality or data integrity'; 'applies to static models'; 'deterministic output' |
| 38 | SARAH | Draft Annex 22 says dynamic models (which keep learning after deployment) should not be used in critical GMP applications. | SUPPORTED | Lines 13-15: 'dynamic models which continuously and automatically learn … should not be used in critical GMP applications' |
| 38 | SARAH | Draft Annex 22 says generative AI including large language models should not be used in critical GMP applications. | SUPPORTED | Lines 20-21, verbatim |
| 38 | SARAH | Draft Annex 22 allows generative AI in non-critical applications with a human in the loop. | SUPPORTED | Lines 21-25: 'If used in non-critical GMP applications … a human-in-the-loop (HITL)' |
| 45 | SARAH | Draft Annex 22 hangs off (is subordinate to / supplements) Annex 11. | SUPPORTED | Scope lines 7-8: 'additional guidance to Annex 11' |
| 70 | SARAH | Under draft Annex 22, a model that retrains itself in a critical application is not permitted. | SUPPORTED | Lines 13-15, as above |
| 76 | SARAH | Draft Annex 22 grades the application as critical or not, and separately rules certain model types out of critical use. | SUPPORTED | Application-level criticality (lines 6-7, 21-22); model-type exclusions (lines 12-21) |
| Line | Speaker | Claim | Verdict | Evidence |
|---|
| 0 | HOST | FDA's AI framework is a draft and never mentions language models. | SUPPORTED | Cover: 'DRAFT GUIDANCE … distributed for comment purposes only'; 'generative', 'language model', 'LLM': 0 hits in the text |
| 16 | SARAH | 'Context of use' and 'model risk' are terms used in FDA's January 2025 AI draft guidance. | SUPPORTED | p.2 'context of use (COU)'; headings 'Step 2: Define the Context of Use', 'Step 3: Assess the AI Model Risk' |
| 19 | SARAH | FDA guidances describe the agency's current thinking. | SUPPORTED | p.2: 'guidances describe the Agency's current thinking on a topic and should be viewed only as recommendations' |
| 20 | HOST | FDA's AI draft guidance is dated January 2025. | SUPPORTED | Cover: 'January 2025' |
| 29 | SARAH | FDA's January 2025 draft concerns AI used to support regulatory decision-making for drugs and biologics. | SUPPORTED | Title: '…to Support Regulatory Decision-Making for Drug and Biological Products' |
| 29 | SARAH | FDA's January 2025 draft sets out a seven-step credibility framework: question of interest, context of use, model risk, then plan, execute and document the credibility evidence, and finally judge whether the model is adequate for its context of use. | SUPPORTED | pp.5-6: Steps 1–7, 'Step 7: Determine the adequacy of the AI model for the COU' |
| 33 | SARAH | FDA's draft says credibility activities should be commensurate with model risk. | SUPPORTED | p.9: 'should be commensurate with the AI model risk and tailored to the specific COU' |
| 33 | SARAH | FDA's draft's worked examples stop at the model-risk grading (step three) and do not describe a sufficient evidence package for a given grade. | SUPPORTED | p.6: 'These two hypothetical examples do not extend beyond step 3' |
| 33 | SARAH | FDA's draft encourages early engagement with the agency. | SUPPORTED | p.10: 'FDA strongly encourages sponsors … to engage early with FDA' |
| 36 | SARAH | FDA's draft is model-type agnostic; it grades the decision rather than the kind of model. | SUPPORTED | p.2: 'does not endorse the use of any specific AI approach or technique'; risk = influence × consequence |
| 40 | SARAH | FDA's draft describes AI models as data-driven and 'self-evolving, capable of autonomously adapting without any human intervention'. | SUPPORTED | p.16: 'data-driven and can be self-evolving (i.e., capable of autonomously adapting without any human intervention)' |
| 40 | SARAH | FDA's draft tells sponsors to anticipate model-directed changes and manage them under lifecycle maintenance. | SUPPORTED | p.16: 'sponsors should anticipate inherent, model-directed changes' |
| 47 | SARAH | FDA's draft asks for a life cycle maintenance plan with performance metrics, a risk-based monitoring frequency, and triggers for retesting the model. | SUPPORTED | p.17: 'model performance metrics, the risk-based frequency for monitoring model performance, and triggers for model retesting' |
| 47 | SARAH | FDA's draft says the detailed life cycle maintenance plan lives in the manufacturing site's pharmaceutical quality system. | SUPPORTED | p.17: 'a component of the manufacturing site's pharmaceutical quality system' |
| 47 | SARAH | FDA's draft says a summary of the life cycle maintenance plan goes into the marketing application for any product- or process-specific model. | SUPPORTED | p.17: 'a summary included in the marketing application for any product or process-specific models' |
| 47 | SARAH | FDA's draft says changes to the model, and manufacturing changes that could affect the model, go through the site's change management system. | SUPPORTED | p.16: 'evaluated by the manufacturer's change management system within their pharmaceutical quality system' |
| 47 | SARAH | FDA's draft says sponsors may use ICH Q12 tools, naming established conditions and comparability protocols. | SUPPORTED | p.17: 'tools outlined in … Q12 … such as established conditions and comparability protocols' |
| 49 | SARAH | FDA's draft does not say which elements of a model would be established conditions or which reporting category a retrain would fall into. | SUPPORTED | p.17: 'may propose model-related elements to be considered established conditions'; no category named |
| 49 | SARAH | The FDA draft says model-related elements may be proposed as Established Conditions. | SUPPORTED | IV.B (p.17): 'Sponsors may propose model-related elements to be considered established conditions, along with a plan to manage changes to these established conditions over the drug product life cycle.' |
| 49 | SARAH | The FDA draft does not say which elements of a model would be Established Conditions. | SUPPORTED | IV.B (p.17) is the only established-conditions passage; it says 'model-related elements' and never enumerates them |
| 49 | SARAH | The FDA draft does not say which reporting category a retraining would fall into. | SUPPORTED | IV.B (p.17): 'the change should be reported to the Agency in accordance with regulatory requirements'; footnote 35 cites 21 CFR 314.70 / 601.12 and two impact factors, but names no category; 'CBE', 'annual report' and 'prior approval supplement' do not appear |
| 63 | SARAH | FDA's January 2025 draft never uses the words 'generative AI' or 'large language model'. | SUPPORTED | 'generative', 'large language', 'language model', 'LLM', 'foundation model': 0 hits |
| 63 | SARAH | FDA's draft says it does not address AI used for operational efficiencies, giving examples: internal workflows, resource allocation, drafting or writing a regulatory submission, provided those uses do not impact patient safety, drug quality, or the reliability of study results. | SUPPORTED | p.3: 'does not address the use of AI models … when used for operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission)…' |
| 67 | HOST | FDA's 2025 draft includes an example of a camera (vision) system that reads vial images to detect a fill-volume deviation. | SUPPORTED | pp.6-7: 'AI-based visual analysis system … fill level in the vials' |
| 68 | SARAH | In FDA's fill-volume example, release testing still measures fill volume on a sample of every batch, so the model is not the sole determinant of release. | SUPPORTED | p.7: 'independent verification of the fill volume is performed on a representative sample for each batch … not be the sole determinant' |
| 68 | SARAH | In FDA's fill-volume example the decision consequence is graded high, the model influence low, and the model risk medium. | SUPPORTED | p.9: 'decision consequence would be high … model influence … low … model risk for this COU is medium' |
| 87 | HOST | The FDA credibility framework assumes fit-for-use data. | SUPPORTED | p.4: 'data used to develop AI models should be fit for use' |
| 90 | SARAH | FDA's draft puts the life cycle maintenance plan inside the site's pharmaceutical quality system. | SUPPORTED | p.17, as claim on line 47 |
| Line | Speaker | Claim | Verdict | Evidence |
|---|
| 10 | SARAH | ICH's reflection paper on advanced pharmaceutical manufacturing was endorsed by the ICH Assembly on 8 October 2025. | SUPPORTED | Running header: 'Endorsed by the ICH Assembly on 8 October 2025' |
| 10 | SARAH | ICH's reflection paper on advanced pharmaceutical manufacturing was published in March 2026. | NOT-IN-CORPUS | Landscape row: 'published Mar 2026'; ICH URL path 2026-03 |
| 10 | SARAH | ICH's reflection paper proposes work and is not itself a guideline. | SUPPORTED | Title: 'Proposed ICH Guideline Work to Facilitate…'; p.9: 'would begin with a new topic proposal' |
| 14 | SARAH | ICH stands for the International Council for Harmonisation. | SUPPORTED | Expanded in the EMA reflection paper and the BioPhorum guidance |
| 31 | SARAH | ICH's reflection paper credits the Points to Consider with the principle that a model's impact guides the extent of regulatory oversight. | SUPPORTED | p.3: 'established the principle that a model's impact … guides the extent of regulatory oversight' |
| 31 | SARAH | ICH's reflection paper says the Points to Consider 'did not explicitly foresee these new types of AI models'. | SUPPORTED | p.4: 'did not explicitly foresee these new types of AI models' |
| 31 | SARAH | ICH's reflection paper says the Points to Consider 'may not be best suited for AI models used as part of a dynamic control strategy and in continuous manufacturing'. | SUPPORTED | p.7, verbatim |
| 58 | SARAH | ICH's reflection paper proposes guideline work on three topics: process modelling (explicitly including AI-based models), continuous process verification (in Q8's sense, not to be confused with continued process verification), and decentralised or distributed manufacturing. | SUPPORTED | p.2: 'process modelling, continuous process verification and decentralised or distributed manufacturing'; p.1 'including artificial intelligence (AI)-based models' |
| 58 | SARAH | ICH's reflection paper's preferred first step is a new ICH guideline on process models. | SUPPORTED | p.9: 'would be a preferred first step' |
| 58 | SARAH | ICH's reflection paper proposes revising the Points to Consider to link model risk to intended use and decision consequence. | SUPPORTED | p.9: 'revising Points to Consider … linking model risk to intended use and decision consequence' |
| 58 | SARAH | ICH's reflection paper lists as an open question what release testing is expected when a model controls the process. | SUPPORTED | p.8: 'expectations for in-process material testing and release testing of product when a model is employed for process control?' |
| 58 | SARAH | ICH's reflection paper asks what lifecycle maintenance means for 'frequently updating process models, for example AI models that learn and self-adjust'. | SUPPORTED | p.8, verbatim |
| 80 | SARAH | ICH's reflection paper is about process models (not about LLM agents). | SUPPORTED | 'LLM', 'large language model', 'agent' do not appear; modelling content is about process models |