Start with the knowledge work your team already does.
Technical references, training packets and administrative records often contain the answer a team needs. Finding the applicable passage is a different task from producing a fluent summary. An evaluation should show which document supports an answer, whether a newer revision changes it and where the available material stops.
Use one approved library and one accountable reviewer. A technical sponsor defines useful questions; IT records the candidate package and network boundary; the reviewer decides whether the resulting answer or draft is acceptable. Keep that division visible throughout the trial.
- Technical references: locate passages, compare revisions and inspect source support.
- Training preparation: draft a study outline from a selected public reference packet, then check each answer.
- Administrative work: prepare a handover draft from synthetic logs with unresolved fields left visible.
A submarine illustrates the connectivity question.
A submarine is an intuitive way to explain why a document workflow may need to continue without an unrestricted cloud connection. Here it is an illustration of a constraint, rather than a customer deployment or an operational procedure.
Local inference is one part of readiness. Installation, model files, document parsing, licensing, updates, diagnostics and recovery can have different dependencies. Test the exact package and chosen task through first launch, restart and document import under the intended network conditions. Record what works, what fails and what still needs a connection.
Inspect a small reference-library example.
The fictional Harbor training-library guide has two revisions. Revision A makes the review-owner field optional and assigns quarterly review to the training coordinator. Revision B makes that field required and assigns monthly review to the reference librarian. Its sections and table rows give an evaluator a precise reference to check.
Try asking who owns the current review, what changed between revisions and where the remote archive is located. The last question has no supplied answer. An acceptable response should preserve that gap rather than invent an address. These are hand-authored reference findings, not recorded Qwerty outputs.
Agree on the supported environment.
Apple Silicon Macs are the planned first platform. A future evaluation will need to record chip, RAM, macOS version, storage, package, model and runtime settings. Establish the supported configuration before planning a wider rollout.
The intended default is local inference, with separate approval of the purpose and data sent to optional external providers. Include file access, proposed writes and denied actions in a future configuration review rather than inferring them from a diagram.
Define acceptance before running the task.
Use questions whose source-backed answers a reviewer can establish independently. Include successful cases, conflicting revisions and absent evidence. Compare ordinary manual work with the candidate workflow using the same material. Count corrections, omissions and time to a reviewed result, keeping failed runs in the record.
Begin the conversation with your document type, intended task, device profile and review owner. Discuss access to nonpublic material through the agreed evaluation process. The public contact path is for describing the task without uploading the library.
- Choose one corpus and a repeatable task set.
- Agree on citation, version and abstention criteria.
- Record the environment and allowed actions.
- Run, inspect sources, correct and retain actual outcomes.
- Decide the next step from accepted work and deployment fit.
Sources and further reading
- Qwerty product and selected-file workflow · Reviewed 2026-10-16
- Qwerty deployment and supported environment · Reviewed 2026-10-16
- Qwerty evaluation methods · Reviewed 2026-10-16