Induced from Abhishek Bachchan’s lecture at the AI in GxP conference (2026-09-15-1426-abhishek-ai-acceptance-criteria-gxp-lecture) and the follow-up conversation with Eric (2026-09-15-1531-abhishek-llm-acceptance-criteria-regulated-industries), 2026-09-15.
Neither CSV nor CSA fits a drifting system
Computer system validation and computer software assurance both assume a deterministic system, a human error-catcher in the loop, and a validated state that persists without further action. An AI system violates all three assumptions: performance drifts as models, prompts, providers, and corpora change, so any validated state has an expiry date. Regulators point the same direction without handing over numeric thresholds: FDA’s SaMD guidance distinguishes locked from adaptive models with predetermined change control plans, the EU GMP Annex 22 draft addresses AI, and roughly four of ten joint FDA/EMA AI principles touch acceptance criteria.
The acceptance-criteria statement names population and threshold
A defensible validation package for an AI system specifies a population (a versioned, change-controlled labeled test set), a risk-justified threshold, the conditions of use, the harness at a named commit, and a deviation route, with run records as the evidence. The conditions of use come from The intended use statement is the master control lever for GxP AI, and the population is tractable only when Bounded inputs make AI validation tractable. Abhishek demonstrated this on a personal RAG project over regulations and warning letters with roughly 100 labels, including drug-versus-device distractors and known-absent probes to catch hallucination. Six acceptance-test buckets cover the space: retrieval, faithfulness, reproducibility, boundary refusal, human-in-the-loop, and guardrails and authority.
Evidence has a shelf life, so ship the dashboard
An inspection-ready package is an interactive dashboard, not a binder: per-call provenance (which model version was requested versus served, prompt hashes), a grounding exception log, human dispositions, threshold rationale, and requalification records. Any trigger, a provider change, a prompt tweak, corpus growth, a chunking change, or simply elapsed time, forces re-measurement. That expiry is AI validation evidence has a shelf life. LLM-as-judge systems agree with human reviewers roughly 76 percent of the time while picking up biases quickly, so they serve as calibrated triage rather than as the judge of record; An LLM judge is a GxP system and needs its own qualification pack is that triage as its own system.
Why this is the regulated-industry instance of external verification
The same principle that makes scientific agents work, verification outside the model, is what regulators demand here: What makes a scientific agent good argues the loop closes outside the model; GxP turns that argument into populations, thresholds, harnesses, and records. The population has to include A golden test set must include unhappy paths, and one way to draft the threshold language is Grounding an LLM in the regulatory corpus yields defensible, novel acceptance criteria. See also Human error, not AI failure, drives GxP warning letters and Validation is data, not an event.
Related
- Home
- AI validation evidence has a shelf life
- An LLM judge is a GxP system and needs its own qualification pack
- A golden test set must include unhappy paths
- Grounding an LLM in the regulatory corpus yields defensible, novel acceptance criteria
- Human error, not AI failure, drives GxP warning letters
- Validation is data, not an event
- The intended use statement is the master control lever for GxP AI
- Bounded inputs make AI validation tractable