An offline AI acceptance checklist should verify the client's promised workflow, not just that a model produces text without internet. Include installation, source processing, output quality, saved files, recovery, and operating ownership.
Write observable criteria with controlled fixtures. This lets the client review concrete outcomes without judging unfamiliar runtime internals or accepting broad capability claims.
Organise checks by the promise they establish
Separate offline availability, task quality, data boundary, hardware fit, recovery, and handover. A pass in one category cannot automatically compensate for a failure in another.
Record exact app, runtime, model, backend, hardware, operating system, source fixtures, and test conditions. Use current runtime documentation as context, then verify the implementation actually delivered.
Review a fictional local lesson tool
Availability criteria can require setup completion, full disconnection, restart, source opening, bounded generation, editing, and presentation. Quality criteria can require preserved facts, source references, checked answer keys, and a useful response when evidence is missing.
Recovery criteria should include interrupted download, corrupt candidate, insufficient storage, cancelled generation, failed extraction, and model unavailability while opening saved material. Data criteria should trace diagnostics, exports, updates, and optional remote paths.
Use harmless or appropriately authorised sources. No benchmark, successful acceptance result, or classroom outcome is claimed in this proposed checklist.
Make evidence and exceptions explicit
For each criterion, name the expected outcome, reviewer, artefact, result, and unresolved limit. Distinguish a blocked test from a failed test and from a case intentionally outside scope.
Define which failures prevent acceptance and which can be accepted with an owner and date. Private-data exposure and lost user work deserve different treatment from minor wording or layout defects.
Include account ownership, reproducible setup, support references, model acquisition, update policy, and independent access to saved outputs. A client should not depend on one developer's laptop to operate the delivered product.
The tradeoff is coverage versus practical review. A concise set tied to real tasks is more useful than a long generic checklist nobody executes. Add cases when the product promise or observed failure gives a reason.
The ChalkLens portfolio account documents a complete local teaching workflow while leaving outcomes unmeasured. This checklist is proposed guidance and does not report acceptance testing of that application.
Questions about offline AI acceptance
Is a disconnected demo enough?
It establishes only the demonstrated path. Verify restart, files, failures, resource fit, and data flow for the actual supported task.
Should the client choose technical mechanisms?
The client should approve outcomes and constraints. Engineering should explain and verify the implementation choices needed to fulfil them.
What should the final review deliver?
Deliver executed criteria, evidence, blocking gaps, accepted limits, owners, and handover artefacts. Bring unclear promises to an offline AI acceptance review.