What can the model see, infer and change, and how can we prove the boundary held?
REFERENCE PROCESS / CONCEPTUAL
01 / ENGINEERING NOTE
Decompose the task
Authorization, exact calculations, entitlement checks and policy evaluation should remain deterministic when the rule is explicit. Use models where language and ambiguity create real value, and bind the result to allowed tools and data. A model should not invent permission by phrasing an action confidently.
Draw the model/tool/policy boundaries before discussing model size.
02 / ENGINEERING NOTE
Authority begins at ingestion
A retrieved paragraph can be outdated, out of scope, or subject to use restrictions. Preserve source identity, publication and ingestion dates, rights, versions and the access context. A citation without an enforceable authorization path is not a governed answer.
Verify source authority independently of linguistic fluency.
03 / ENGINEERING NOTE
Evaluation must reflect failure cost
General benchmarks rarely exercise the workflow’s important exceptions. Build examples for missing sources, adversarial instructions, stale content, ambiguous calculations and denied actions. Version expected outputs and record evidence for the failures that matter.
Test for safe refusal as well as successful completion.
04 / ENGINEERING NOTE
Make uncertainty inspectable
Show the sources and tool results used, where the model expressed uncertainty, and what requires human confirmation. For consequential actions, require explicit review and an audit trail. The experience should make unsupported certainty difficult.
A user should be able to challenge a result without seeing an opaque score.
These notes describe design reasoning, not a product certification or client deployment.
← Back to all insights