Module 5: Research, Retrieval, Standards, and Evidence Engineering

Learning Objectives

  • Use research agents and retrieval tools to locate authoritative engineering evidence without treating search summaries as authority.
  • Distinguish source discovery, source verification, evidence extraction, synthesis, and engineering decision-making as separate stages.
  • Track edition, revision, jurisdiction, provenance, access date, and verification status for material claims.
  • Evaluate citation correctness, citation completeness, source sufficiency, and conflicts between authoritative records.
  • Design bounded research workflows that stop or escalate when primary evidence is unavailable, proprietary, superseded, or contradictory.

Primary Engineering Source

A primary engineering source is an original authoritative document issued by the responsible standards body, government agency, professional institution, project owner, or contract authority, such as a code, standard, official manual, approved specification, or formal issuance.

Source Verification Status

Source verification status records whether a material engineering claim is confirmed from a primary source, supported only by a secondary source, ambiguous because sources conflict, or unverified because the authoritative document is unavailable.

Research Hierarchy for Civil Engineering

Prefer evidence in this order when the task depends on authority:

  1. Governing law, regulation, code, or official agency issuance.
  2. Official standard or technical manual from the issuing organization.
  3. Project contract documents, approved drawings, specifications, and revisions.
  4. Peer-reviewed research or recognized technical guidance.
  5. Reputable secondary interpretation.
  6. General web commentary only for orientation, never as the sole basis of a consequential compliance claim.

The hierarchy must still be interpreted in the context of jurisdiction, adoption, project contract hierarchy, and effective date.

Engineering Research Verification Workflow

A source-first path for standards, codes, and technical guidance.

Follow the path from a research question to a verified engineering claim; inaccessible sources remain evidence gaps.

Engineering Research Verification WorkflowA source-first path for standards, codes, and technical guidance.. Define engineering research question → Locate primary or official sources; Locate primary or official sources → Check edition, revision, jurisdiction, and status; Check edition, revision, jurisdiction, and status → Material claim directly verified?; Material claim directly verified? — Yes → Record verified claim with source traceability; Material claim directly verified? — No → Record evidence gap or secondary-only status; Record verified claim with source traceability → Use evidence in reviewed engineering work; Record evidence gap or secondary-only status → Use evidence in reviewed engineering work

Define engineering research question → Locate primary or official sources; Locate primary or official sources → Check edition, revision, jurisdiction, and status; Check edition, revision, jurisdiction, and status → Material claim directly verified?; Material claim directly verified? — Yes → Record verified claim with source traceability; Material claim directly verified? — No → Record evidence gap or secondary-only status; Record verified claim with source traceability → Use evidence in reviewed engineering work; Record evidence gap or secondary-only status → Use evidence in reviewed engineering work

  • Define engineering research question: terminator
  • Locate primary or official sources: process
  • Check edition, revision, jurisdiction, and status: process
  • Material claim directly verified?: decision
  • Record evidence gap or secondary-only status: document
  • Record verified claim with source traceability: document
  • Use evidence in reviewed engineering work: terminator

Research Prompt Pattern

Research the question using primary or official sources first.
For every material requirement, provide:
- issuing organization;
- document title;
- edition or revision date;
- current/superseded status where known;
- clause/section only if verified in the source;
- direct source link where legally accessible;
- a short explanation of applicability;
- source verification status;
- any conflict or uncertainty between sources.
Do not invent section numbers, quotations, or requirements.
Engineering Source Authority Ladder

A hierarchy for selecting evidence when authority matters.

Read downward from the strongest authority; project applicability and adoption still need to be established.

1
Law, regulation, or official issuance
2
Official code, standard, or agency manual
3
Approved project and contract documents
4
Peer-reviewed or recognized technical guidance
5
Secondary commentary for orientation only

Current AI Research Workflows — September 24, 2026

Current research-capable systems include ChatGPT Deep Research, OpenAI agent workflows that can coordinate tools and files, Google Deep Research and Deep Research Max, and Claude workflows using current model families, connected sources, and scientific/research workbenches. Their engineering value is in discovery, synthesis, evidence organization, source comparison, and gap detection rather than replacing authoritative-source review.

Research systems are especially useful for:

  • mapping a topic before detailed review;
  • running repeatable source-restricted searches with an evidence log;
  • finding current government or standards-body publications;
  • checking publication and revision histories;
  • identifying conflicting or superseded guidance;
  • building evidence matrices with source links;
  • locating gaps that require licensed copies, project records, or expert interpretation.
Research Evidence Status Map

Every material claim should carry an explicit verification state.

These status labels prevent a discovered citation from being mistaken for a verified engineering requirement.

Discovered
Verified
Secondary-only
Conflicting
Inaccessible
Superseded

Research Agent Evidence Contract

For consequential engineering research, require the agent to distinguish:

  • discovered: the source was located but its material claim has not yet been checked;
  • verified: the material claim was checked directly against an authoritative source;
  • secondary-only: only commentary or a secondary interpretation is available;
  • conflicting: authoritative sources disagree or applicability is unclear;
  • inaccessible: the governing source cannot be lawfully or technically inspected;
  • superseded: a newer edition, amendment, issuance, or adoption controls.

The final research artifact should preserve source URL or document ID, issuer, edition/date, access date, claim supported, and verification status.

Citation Versus Requirement Verification

Citation existence and requirement correctness are separate checks.

A real document link is not enough; the actual requirement, edition, scope, and applicability must match the claim.

Citation check
  • Document exists
  • Issuer is authentic
  • Link reaches intended source
Requirement check
  • Edition is applicable
  • Clause supports the claim
  • Definitions and exceptions checked
  • Jurisdiction and adoption confirmed

Verify the Requirement, Not Just the Citation

A citation can point to a real document while the claimed requirement is still wrong. Confirm wording, edition, scope, definitions, exceptions, tables, figures, adoption status, and referenced sections in the original source before relying on a compliance statement.

Proprietary Standards and Inaccessible Sources

Many engineering standards are copyrighted or paywalled. If the AI cannot lawfully access the full current text, it should not pretend to verify a clause. Use AI to organize the review, then consult a licensed copy or authorized source.

Version, Jurisdiction, and Status Control

Record at minimum:

  • standard title and designation;
  • edition/year;
  • amendments, addenda, or errata;
  • current, superseded, draft, or historical status where relevant;
  • jurisdiction or adopting authority;
  • project specification or contract overrides;
  • source URL or controlled document identifier;
  • date the source was checked.

A technically accurate requirement from the wrong edition can still be inapplicable.

Evidence Maintenance Loop

Current engineering research is a controlled update process.

Research agents should revisit the evidence matrix when new or conflicting sources appear.

Traceable evidence matrix
Discover source
Classify authority
Verify claim
Record status
Recheck when evidence changes
↺ repeat when checks fail

Worked Use Cases and Reusable Prompts

The companion Worked Examples includes standards-landscape research, citation existence checks, edition comparison, government-domain research, literature matrices, stale-source detection, evidence-gap reporting, and standards research logs.

Build an Evidence Graph, Not a Citation List

For consequential research, connect each material engineering claim to the source that supports it. Record the source identity, edition or revision, exact location when available, access status, and verification state. A bibliography proves that sources were collected; an evidence graph shows which source actually supports which conclusion.

Separate Retrieval from Interpretation

A robust research agent should proceed through controlled stages:

  1. discover candidate sources;
  2. classify source authority and currency;
  3. retrieve the primary material when legally and technically available;
  4. extract only the evidence relevant to the engineering question;
  5. verify the citation against the original source;
  6. identify conflicts, gaps, or superseded requirements;
  7. synthesize findings with verification status visible;
  8. stop for human judgment before a requirement enters a consequential engineering decision.

This separation reduces the chance that a search snippet or secondary summary is silently promoted into a governing requirement.

Audit Citation Quality Along Four Dimensions

  • Correctness: the cited source actually supports the stated claim.
  • Completeness: material claims that require evidence are cited.
  • Sufficiency: the authority and detail of the source are strong enough for the claim being made.
  • Currency: the edition, revision, jurisdiction, and effective date are appropriate for the project.

A response can contain real links and still fail one or more of these dimensions.

Research Agents Need Stop Conditions

Stop or escalate when the authoritative source cannot be accessed, the applicable edition is uncertain, two primary sources conflict, jurisdiction is unclear, a proprietary standard is being reconstructed from secondary commentary, or the requested conclusion exceeds what the evidence can support. A well-designed research workflow is allowed to return not verified.

Key Takeaways
  • AI can accelerate engineering research, but authority comes from the governing source, not the model.
  • Edition, status, jurisdiction, revision, and project applicability must be controlled explicitly.
  • Real citations can still support false claims; verify the actual requirement in the source.
  • Proprietary or inaccessible standards must be treated as unverified until checked from an authorized copy.
  • Research outputs should preserve evidence links, verification status, uncertainty, and conflicting guidance for human review.

References

  • Deep research in ChatGPT — OpenAI. Official guidance on research across the public web, uploaded files, selected websites, and enabled apps with cited reports.

  • Introducing the Agents API — OpenAI, 2026-09-10. Current agent infrastructure for long-running, tool-using workflows with files, code execution, context management, and subagents.

  • Deep Research Max — Google, 2026-04-21. Official description of long-horizon research agents, connected sources, MCP support, and native analytical outputs.

  • Introducing Claude Opus 5.5 — Anthropic, 2026-09-22. Current Anthropic model release relevant to long-running agentic and professional work.

  • Claude Science, an AI workbench for scientists — Anthropic, 2026-06-30. Official description of an auditable workbench with tools, agents, citations, calculations, and reproducible artifacts.

  • Artificial Intelligence and engineering responsibility — ASCE. Professional responsibility remains with the engineer when AI is used to discover or synthesize technical requirements.

  • Model optimization — OpenAI. Evaluation-driven methodology relevant to testing research outputs and citation behavior.

  • ARIA Evaluation Planning Manual — NIST, 2026-09-18. Holistic AI-evaluation guidance combining model testing, red teaming, and user testing.