Chapter 6 — Professional Practice Integration

---

6.1 Introduction

Using agents in professional work raises questions about supervision, competence, examination, documentation and reliance. The Alberta engineering setting attaches concrete obligations to these questions.

Scope of the regulatory analysis. The mapping retains the source editions and interpretation used in the thesis as of 2026-07-02. This analysis reassesses the architectural argument; it does not establish whether those regulatory sources are current, amend the firm's practice management plan or certify legal compliance. The mapping is the author's interpretation, not a determination or endorsement by APEGA. Appendix C sets out in detail how the review topics relate to the proposed support.

---

6.2 Regulatory Framework

6.2.1 Governing Legislation and Standards

The analysis situates the work under Alberta's Engineering and Geoscience Professions Act, its General Regulation and three APEGA practice standards:

• Relying on the Work of Others and Outsourcing, May 2021, v4.0 [APEGA_RWO2021];
• Authenticating Professional Work Products, November 2024, v8.6 [APEGA_Auth];
• Professional Practice Management Plan, November 2022, v1.1 [APEGA_PPMP].

The cited editions supply the section references below. Their applicability, and whether they are current, must be established for actual work.

The firm's Professional Engineering with Agentic AI in Regulated Practice is maintained in its quality management system [Chirality_PE].

6.2.2 Professional Work Products and Reliance

In the cited practice context, a professional work product (PWP) concerns technical information on which others rely to make decisions or take action. Its character is not created by a repository status field. Nor does an unauthenticated work product become a PWP only when someone authenticates it.

The relevant distinctions are the work's content, purpose, intended reliance, state of examination and applicable professional acts. A draft can contain technical information that matters. A completed software change has its own subject of acceptance and need not be an authenticated engineering work product.

Git integration, an owner accepting a development slice and professional authentication are different acts, and the subject and authority of each must be identified separately.

6.2.3 Applicability to AI-Assisted Work

The dated source basis includes APEGA's Guidance for Registrants Regarding the Use of Artificial Intelligence Tools, updated March 23, 2026 [APEGA_AI2026]. The interpretation reads that guidance as placing responsibility for appropriate use and examination of AI results with the registrant, while comparing AI results to those of other technical tools.

The thesis proposed using the supervision and review topics from the reliance standard to structure AI-assisted practice. That proposal remains useful as a design inquiry, but it does not settle whether an AI agent is an “other” in the standard's legal meaning. Being able to map a mechanism to a professional concern establishes neither direct legal applicability nor sufficient performance of the obligation.

---

6.3 The Central Argument: AI Agents as "Others"

6.3.1 Structure of the Argument

The argument is:

1. Professional use of agent-produced work raises concrete questions about scope, competence, evidence, examination and responsibility.
2. The cited supervision and review topics offer a disciplined way to examine those questions.
3. Agent instructions, tools and records can support some of the required activities, subject to their actual operation and limitations.
4. The architecture cannot establish that the professional performed those activities adequately or that the arrangement satisfies applicable law.

This argument supports a candidate practice arrangement without depending on a settled classification of the agent as a worker or tool.

6.3.2 Direct Application and Analogy: An Interpretive Limit

The “others” interpretation offers a conservative candidate position. The alternative treats AI as a technical tool whose results the professional must be competent to assess. Neither interpretation can be settled simply by calling the system an agent or by observing that it performs complicated work.

Classifying AI as a tool does not imply that casual review is adequate, and an analogy to a worker does not prove that extra records or approvals improve protection. The appropriate examination depends on the work's consequences, the kinds of error that are possible and the applicable professional requirements.

Agents can interpret requirements, compare alternatives, criticize designs and take authorized action. In the thesis's terminology, this is sophisticated reckoning. Judgment is the person's situated, committed and world-involving answerability for the work. The distinction does not restrict agents to clerical production, nor does it make their legal classification follow from their capability.

6.3.3 The Technology Provider Distinction

A model supplier, the host application, the agent's assignment and the professional work are separate things to examine. A supplier's claims about capability do not establish that a particular output is suitable for reliance. Likewise, an instruction framework does not make the supplier's component safe by definition.

In practice, the agent system includes the model, instructions, context, tools, permissions and execution environment, and investigating a failure may require examining any of these. A model update can change behaviour; a tool can return wrong units; a permission defect can allow an unintended operation.

The thesis allocates accountable professional commitment to the person using the work. It makes no general legal determination here about the liabilities of suppliers or firms, and the architecture's responsibility model cannot settle those questions.

---

6.4 Direct Supervision and Control

6.4.1 The APEGA Standard

The mapping uses §3.1.1 of the cited reliance standard to examine active involvement, responsibility in decision making and documentation. These topics direct attention to what the professional does throughout the work, but they do not demonstrate that a particular number of agents, levels or approval stops is necessary.

Chirality contributes proposed support for that conduct. Actual supervision cannot be inferred from a passed validator, an available dashboard or an instruction telling an agent to ask when it is uncertain.

6.4.2 Active Involvement (APEGA §3.1.1.1)

Five topics organize this inquiry:

Agent 0 maintains continuity; managers and bounded contributors are used where they serve the work. Roles do not grant permissions. Write access follows the assignment and host limits, and TASK does not delegate in the repository's chosen role framework. That arrangement is one current implementation, not something deduced from the professional topics above.

6.4.3 Responsibility in Decision Making (APEGA §3.1.1.2)

In the cited May 2021 standard, §3.1.1.2 assigns the licensed professional responsibility for all technical engineering or geoscience decisions related to the work. It also requires consideration and documentation of relevant issues, technical direction with attention to applicable requirements, and availability to answer questions and review and approve decisions made by those doing the work [APEGA_RWO2021] (p. 11). The source does not limit these obligations to a repository's reserved decisions.

The repository separately permits standing authorization for investigation, implementation and repair. Ordinary action need not return to the owner merely because a procedural stage has changed, but changes to reserved commitments, acceptance criteria, reliance purposes or risk decisions require the appropriate person. These are repository arrangements, not a narrower interpretation of the cited professional obligation. Where that obligation applies, operational selections and standing authorization remain subject to it; calling a choice ordinary implementation does not remove professional responsibility for a technical engineering or geoscience decision.

An agent can present a choice and its consequences, resolve an implementation question within actual authority, or apply an already-authorized decision without pretending to exercise the person's judgment. The required professional involvement must still be established for the actual practice context.

The decision changes the artifact it governs. Its reason and scope are preserved where the work needs them, using the existing change record when that record is adequate. Removing an administrative ledger does not remove an actual professional documentation requirement.

6.4.4 Documentation of Due Diligence

Documentation should let a relevant reader reconstruct what was directed, examined and relied upon. The amount and form depend on the practice context and applicable requirements. Current source documents, identified outputs, captured conversation and version history may supply that evidence; a special record is justified where these are insufficient.

Git preserves committed content and its changes, but it does not automatically capture every conversation, action, input, human decision or external event. Nor does a versioned instruction prove that it was followed. A claim about practice must state what its evidence actually covers.

The person can perform an authorized act through a host that captures the result. A machine-written record of that act is not an acceptance that originated with the machine. Conversely, a generated statement that approval occurred cannot create an approval that never happened.

---

6.5 Thorough Review and Authentication

6.5.1 Applicability

The mapping uses §3.1.2 of the cited reliance standard to organize the examination that comes before professional reliance. Review in this sense is more than format checking: it examines whether the work and its grounds support the intended purpose.

No mandatory lifecycle for claims follows. Cited, reviewed and authenticated identify different relationships; they are not increasing truth values. The appropriate review must consider material claims that labels or dependency metadata may have missed.

6.5.2 Reliability, Accuracy, and Validity (APEGA §3.1.2.1)

Review addresses the following subjects in relation to the work's intended use. Who carries out the review, and when, depends on the required competence, consequences and grounds.

An agent can discover discrepancies, prepare a focused comparison and carry repairs through the authorized work. Independent scrutiny is particularly valuable where contributors may share an undetected premise. It can involve another competent person, another agent or a different technical method as appropriate to the claim; an agent reviewer does not substitute for a required professional reviewer.

6.5.3 Adherence to Regulatory Requirements (APEGA §3.1.2.2)

The historical mapping included legislation, practice standards and project-specific approvals or permits. Current work must identify its applicable requirements and authoritative source editions. An instruction to obey them is behavioural guidance, not a compliance check. A checker can establish a bounded property only if its rule correctly represents the applicable requirement and its coverage is adequate.

6.5.4 Adherence to Quality Control and Assurance (APEGA §3.1.2.3)

The mapping also included the practice management plan and the clarity, consistency and completeness of work. No document kit can ensure these properties by its presence. A short current basis can be clearer than a complete set of contradictory templates.

A changed repository procedure does not automatically amend an external practice obligation. The responsible person must address any applicable practice-system requirement in its actual home. This chapter neither adds such an obligation nor declares it discharged.

6.5.5 Authentication

The cited authentication discussion concerns a professional act relating a person to identified work and responsibility for its use. This framework reserves that act to the appropriate person; a PR merge, a successful analysis or a generated receipt does not perform it.

Content addressing can identify the bytes associated with the act, but reliable attribution, scope and purpose remain separately necessary. A fingerprint alone proves neither who approved the work nor whether the examination was adequate. These limits apply whether the person types the record or an authorized host interaction captures it.

A later change leaves the historical act attached to its original subject. Whether a prior acceptance remains applicable to changed work requires assessing the affected content and reliance conditions. A universal rule that any repository edit voids all earlier acceptance is not warranted.

Authentication does not create knowledge, establish truth or exhaust judgment. The person's committed engagement also shapes purpose, framing, interpretation and action before this particular act occurs.

---

6.6 Professional Obligations

This section states the framework's adopted commitments. Their legal application remains subject to the historical-source and interpretive limits in §6.1; software does not guarantee their fulfilment.

6.6.1 Public Welfare as First Constraint

The framework puts public safety and the protection of people, property and the environment ahead of schedule or convenience. Material hazards require appropriate examination and decision. A severity label can help direct attention, but it cannot guarantee that hazards were found or effectively controlled.

Delay, excessive paperwork and the diversion of expert attention also have consequences, and their costs belong in the design assessment. Removing a low-value check can improve practice when it frees attention for an important failure mode; removing a numerical oracle can worsen it when wrong results would otherwise remain plausible.

6.6.2 Responsible Charge Remains Human

Chirality places accountable professional commitment with the person. Agents may perform extensive reckoning and act within the authority granted to them, but they do not thereby acquire the person's responsibility or authority to authenticate. This is the framework's allocation for the systems examined here, not a metaphysical proof about every possible future machine.

Judgment is exercised throughout the work. Reducing it to the final signature would make an unattended approval ritual appear sufficient. A person's decision needs engagement with what the work means and what they will stand behind; a record can serve as evidence of an act but cannot guarantee that engagement.

6.6.3 Competence Includes Tool Competence

The professional needs a defensible basis for assessing the tools and work used for the intended purpose. This includes knowing what a check establishes, where a model or calculation may fail and when further expertise is needed, but it does not mean that every computation must be repeated manually.

Useful evidence can include analytical solutions, validated benchmarks, independent methods, input inspection and observed behaviour. Competence cannot be inferred from having a checklist, tool catalog or instruction file. The architecture can make evidence accessible, but the person must be able to understand its relevance and limits.

6.6.4 Evidence Over Plausibility

Material claims need suitable grounds. An extracted number must not quietly become an inferred estimate; an accepted assumption must not be presented as an observation. A source citation needs to support the actual claim, and a passing test needs a stated scope.

This is a commitment implemented through instructions, useful data structures and focused checks. It is not proof that invention or omission cannot occur.

6.6.5 Hierarchy of Authority

Authority, evidential strength and task responsibility are different orders. An applicable law or contract constrains action; it is not a scientific observation. Strong evidence can show that a chosen model is wrong without authorizing an agent to change a reserved commitment. Professional judgment interprets applicability and acts within constraints; it is not a source of evidence ranked last, below every published document.

Current instructions and assignments identify where decisions belong. Roles help organize contributions, but their names or numerical types do not create professional authority.

---

6.7 Quality Control and Assurance

6.7.1 Architectural Controls

A proposed control should identify the failure it addresses, the evidence it produces and its cost. Schema validation can reject malformed inputs, path containment can refuse an escaping write, and a numerical oracle can detect a particular class of calculation error. Each has a bounded claim.

A candidate arrangement of controls needs to be evaluated in actual practice. Adequate support for professional reliance concerns the work, examination and engagement it enables. Architectural coherence supplies a rationale to examine, not a finding that obligations have been discharged.

Current assurance selects checks according to consequence, detectability, overlap and cost. Reuse applicable evidence and repeat a check when a relevant change or unresolved concern warrants it. A requirement to demonstrate a specific property still applies even when the general process becomes lighter.

6.7.2 Instruction Governance as Release Engineering

Instruction changes can alter behaviour and deserve to be versioned and reviewable. The depth of examination depends on the affected behaviour. A permission change warrants different scrutiny from a clarified entry point.

Instructions must also agree with their enforcement. Removing a read-only sentence while a resolver still discards authorized write targets does not change the effective policy. Conversely, broader tool availability must not silently become broader authorization. The relevant consumer and enforcement path belong in the same change.

6.7.3 The Deterministic and Probabilistic Boundary

Use reliable tools for operations they perform well: resolving identities, comparing revisions, applying a structured operation or evaluating a defined formula. Agents can then investigate, interpret, propose and coordinate around those results.

Deterministic does not mean correct. A repeatable calculation can use a wrong formula, precision rule, unit conversion or design basis. The strength of its evidence depends on validated behaviour and applicability. Agents may also make useful observations about a deterministic tool's failure.

The division of labour is open to revision. A reusable tool is worthwhile where it improves reliability or effort enough to justify building and maintaining it. That an algorithmic solution is possible does not require a new tool for every one-off action.

6.7.4 Technology Provider Due Diligence

The provider and model affect the system's behaviour. Useful evaluation asks whether the available capability, failure modes and execution conditions fit the work, including privacy, permission and recovery needs. Supplier claims and general benchmarks cannot settle those questions for a particular application.

As models improve, a control added for an earlier limitation can become unnecessary overhead. Reassess the control against observed behaviour and remaining risk. Stronger output does not by itself transfer accountable judgment or remove the need to inspect results that matter.

6.7.5 Evidence and Auditability

A usable evidence arrangement makes relevant inputs, results, changes and acts retrievable by an appropriate reader, and it states what was captured and what was not. Source history, product logs and existing review records may be sufficient; routine copies of the same information can obscure inspection rather than improve it.

Three questions guide assessment: can a reviewer identify what supports the claim, can a participant recover the state needed to continue, and can an appropriate reader reconstruct an act that matters? Answering them requires evidence from actual use, and an exhaustive-looking register is not a substitute for it.

---

6.8 The Human-AI Contract

The contract here is the framework's allocation of responsibility, not a new grant of product permissions or a claim that every instruction is mechanically enforceable.

6.8.1 Human Responsibilities

The person establishes or accepts the work's purpose, material commitments, limits and reliance conditions. They provide or arrange the competence needed for decisions that matter, determine reserved risk acceptance and perform any required professional act. They can authorize agents to continue ordinary work without repeated approval.

Agent 0 supports continuity with that person but does not replace their judgment. Meaningful engagement includes framing and revising the work, not merely approving an output prepared elsewhere.

6.8.2 AI Agent Permitted Actions

Within the actual assignment and host permissions, agents can investigate, extract, infer, plan, evaluate, criticize, propose, implement, check, repair and integrate. They can select ordinary implementation details and apply decisions already authorized. These activities may involve substantial initiative.

Agents can make operational selections within their assignment and adopted policy. Their capability supports extensive initiative, but authority for reserved commitments and accountable professional reliance remains with the person.

6.8.3 AI Agent Prohibitions

The framework prohibits agents from inventing a human act of acceptance, claiming professional authority they do not possess, concealing a material conflict, presenting fabricated evidence as observation or exceeding the assignment and host permissions.

Enforcement must be described accurately. A host can refuse an out-of-scope operation; an instruction requests behaviour; review can detect some violations. None alone guarantees all the others. The prohibition and the mechanism that attempts to enforce it are separate claims.

---

6.9 Limitations of the Regulatory Mapping

6.9.1 The Mapping Is an Interpretation, Not a Regulatory Ruling

The mapping is an argument by the author using the cited historical sources. It has not been established here as endorsed by APEGA or accepted by a court, tribunal or professional reviewer. Its architectural examples are candidates for support, not certificates of compliance.

6.9.2 Jurisdiction Specificity

The sources concern Alberta engineering practice. Similar practical questions may arise elsewhere, but their legal meaning and required conduct cannot be transferred merely by replacing the regulator's name. This chapter establishes no claim of compliance across jurisdictions.

6.9.3 The "Relying on the Work of Others" and "Using a Tool" Distinction

Both classifications deserve consideration. The proposed “others” argument does not follow necessarily from the text's emphasis on professional conduct. Nor does a more demanding internal practice policy prove compliance with a different legal interpretation. The support that discoverable evidence and controlled operations offer can remain useful while that classification is unresolved.

6.9.4 The Architecture Is One Candidate, Not a Necessary or Sufficient Condition

The thesis has not shown that Chirality is required for competent practice. Other arrangements can preserve evidence and responsibility. Nor has it shown that installing Chirality is sufficient: inadequate examination, wrong assumptions, misunderstood tools and mechanical approval can still occur.

The appropriate test is how an actual arrangement performs, including its failures and the attention, delay and complexity it imposes.

6.9.5 Evolving Regulatory Context

The section references and interpretive mapping retain their 2026-07-02 source basis. This revision has not rechecked whether they are current. A current professional application requires a separate examination of the applicable sources and practice requirements. Repository development experience can inform that examination but cannot determine its legal outcome.

---

6.10 Summary

The professional-practice analysis offers concrete questions for designing agent-assisted work: what was directed, what supports the result, what was examined, which limits remain and who stands behind reliance.

Chirality proposes ways to make those questions answerable. Continuing ownership, discoverable current artifacts, useful checks and attributable acts can support practice without a fixed hierarchy or a routine chain of records. Their effectiveness remains an empirical question. Instructions, content hashes and passed tests each establish only bounded facts.

The enduring commitment is to distinguish extensive agent reckoning from the person's situated judgment throughout the work. Greater capability can change which supervision and verification arrangements are appropriate without itself transferring that accountable engagement. The revised mapping preserves this commitment while withdrawing the claim that one architecture guarantees, or is necessary for, professional compliance.