Induced from Abhishek Bachchan’s acceptance-criteria lecture at the KENX AI in GxP East 2026 conference (2026-09-15-1426-abhishek-ai-acceptance-criteria-gxp-lecture), 2026-09-15.
Triggers that expire a validated state
A model or provider change, a prompt tweak, corpus growth, retrieval or chunking parameter changes, a context-of-use shift, or simply elapsed time. Any one of them forces re-measurement, which makes validation a continuous loop rather than the one-pass V model that assumes a validated state persists without action. This is the operational form of Validation is data, not an event, and it is why Validating an AI system in a regulated industry means validating a measurement system treats every validated state as having an expiry date.
Evidence is a live database, not a PDF
The inspection-ready package is a dashboard: labeled test sets with rationales, dated run records, per-call provenance (requested versus served model, prompt hashes, token counts), grounding exception logs, stored outputs, human dispositions, threshold rationale, and requalification records. Screenshots age; queries against the record do not. Observability and evaluation answer different questions about an AI system supplies the two streams that dashboard has to keep current, and Centralize evaluation when every AI pairing carries its own KPIs is how those streams stay inspectable when every pairing ages on its own clock.
Related
- AI Validation and Acceptance Criteria MOC
- Validation is data, not an event
- Validating an AI system in a regulated industry means validating a measurement system
- Observability and evaluation answer different questions about an AI system
- Centralize evaluation when every AI pairing carries its own KPIs