Choose bounded tasks
Useful starting points include searching approved project documents, drafting model descriptions and explaining the structure of a results table. Avoid treating fluent text as evidence that a calculation, standard or citation is correct.
Build retrieval around provenance
- Use an approved document collection with versions, access permissions and publication dates.
- Retrieve passages relevant to the question and retain their page or section references.
- Require the answer to identify its evidence and state when the collection is insufficient.
- Verify the cited passage supports the conclusion, rather than only mentioning the topic.
Keep calculations reproducible
Run hydraulic calculations and data transformations in controlled tools with logged inputs. A language model can help prepare an analysis, but the numerical result should be reproducible outside the conversation. Have a qualified person review assumptions and conclusions.
Evaluate before routine use
Create representative questions, including ambiguous and unanswerable cases. Check unsupported claims, retrieval failures and permission boundaries. NIST’s AI Risk Management Framework provides a general risk-management reference; it is not a water-design standard.
Choose a task with a verifiable output
Useful tasks include locating a passage in approved guidance, drafting a model-description paragraph from checked inputs or helping structure a data-processing script. Each has an output that can be reviewed against evidence or executed independently. Asking for an unsupported design capacity combines assumptions and calculations in a way that fluent text can conceal.
Define the document collection, version and project context. A historical standard or an unrelated utility guideline can look authoritative while being inapplicable. The system should be able to say that the approved collection does not contain enough information.
Build a source-grounded answer format
For a project question, require the answer to separate source facts, calculated results, assumptions and unresolved questions. Attach a document title, version and page or section to the source-dependent statements. A link to a long document without a supporting location is difficult to review.
Check that the cited passage actually supports the claim and that the answer has not combined incompatible requirements from different jurisdictions. Where two sources disagree, present the conflict and its scope instead of selecting whichever statement sounds more convenient.
Retrieved documents are evidence, not instructions for the assistant to follow. Text inside a document or webpage can contain irrelevant or malicious directions. Keep system behaviour, access permissions and approved tools controlled outside the retrieved content.
Keep engineering calculations outside unverified prose
A language model can propose a formula or draft a script, but the numerical result should be produced in a reproducible calculation with named inputs and units. Check dimensional consistency, boundary conditions and an independent sample. A correct arithmetic result can still answer the wrong engineering question.
For example, a reservoir calculation should distinguish total volume from usable volume and a temporary deficit from an ongoing replenishment shortfall. Requiring the assistant to show those assumptions is more useful than asking it to “be accurate” without a structured task. Use the storage guide as the engineering basis.
Evaluate retrieval and answering separately
| Test case | What a useful system should do |
|---|---|
| Answer clearly supported | Return the relevant statement with a precise source location. |
| Ambiguous project context | Identify the missing context before applying a requirement. |
| No supporting evidence | State the evidence gap instead of inventing a citation. |
| Outdated or conflicting versions | Show the version issue and preserve the distinction. |
| Restricted source | Respect access controls even if another user previously queried it. |
Measure unsupported claims, citation accuracy and missed evidence, not only whether the response sounds helpful. Use known reference answers and reviewer notes. Keep a set of difficult cases that includes missing documents and questions with no answer in the collection.
Integrate it into a reviewable workflow
Retain the source snapshot, relevant instructions, output and material human corrections for work that influences a project record. Define which uses need technical review and who owns the final decision. Do not allow a draft answer to acquire approval status merely because it is copied into a report.
Use approved systems for confidential project information. The Atlas does not provide a document-upload assistant or connect to private utility data. Its Python guide demonstrates how to keep calculations reproducible, while the cited NIST material provides general AI risk-management context.
Sources & further reading
- AI Risk Management Framework ↗National Institute of Standards and Technology · 2023 framework; 2024 generative AI profile
External source · Checked 24 September 2026
Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.