# Chapter 9 — Discussion --- ## 9.1 Introduction Root is both a means of producing agent-assisted work and a subject of that work. Participants use its instructions, methods and tools to examine and change those same surfaces. This gives the case its generative character and its main risk: the environment can improve itself, but it can also reproduce its own assumptions across producers, tools and reviewers. The strongest case for Root is not that capable agents cannot work without a framework. They often can, and ordinary conversation, files and tools may be enough. Its case is that repeated undertakings expose relationships whose neglect becomes expensive: lost intent, hidden assumptions, disconnected results and unclear authority. Shared means are justified where they preserve useful learning about those relationships at less cost than repeatedly rebuilding it. The four perspectives help expose that risk. A label may be mistaken for an act, a query for an assessment, a record for understanding, or a completed contributor for an integrated result. The framework does not prevent those confusions merely by naming them. Its adequacy must be examined on actual connecting paths. --- ## 9.2 Limitations ### 9.2.1 No Controlled Empirical Validation Root has not been assessed in a controlled comparison of development arrangements or in a controlled study with practising engineers. The repository demonstrates substantial agent-assisted production, examination and repair. It does not establish that Chirality: - improves detection of consequential errors compared with other workable arrangements; - reduces total effort for a comparable accepted result; - improves calibration of professional confidence; - transfers to another practitioner without its author's continuing help; - meets professional regulatory obligations in a particular deployment. The production evidence is bounded by changing agents and suppliers, different work start times, generated material in some diffs and the owner's continued steering. PR, line and file counts describe activity or size rather than comparable completed work. [Appendix E](appendix_e_development_history.md) retains the owner's reports and the development record; neither isolates a causal effect or supplies a productivity multiplier. ### 9.2.2 Regulatory Interpretation, Not Endorsement The APEGA mapping is an interpretation of professional supervision, review and reliance [CITE:APEGA_RWO2021]. Describing an agent as a contributor or a tool does not discharge the professional's duties. An implementation can preserve identified inputs, actions and decisions. It cannot establish that a professional was competent, performed sufficient review, or made a valid professional authentication merely because a record exists. Application acceptance, approval of repository work and professional authentication must remain distinct. The appropriate interpretation and practice requirements need confirmation for the actual jurisdiction, work and date of use. ### 9.2.3 Jurisdiction-Specific Regulatory Framework The detailed regulatory discussion is specific to its Alberta basis. Other jurisdictions and disciplines have their own definitions of competence, supervision, review, authentication and responsibility. Similar language is not evidence that requirements or legal consequences are interchangeable. The portable research question is whether the application makes relevant work, grounds and acts inspectable. The legal adequacy of a deployment requires a separate, current assessment. The four philosophical questions are not a regulatory checklist that certifies its answer. ### 9.2.4 Instruction-Level Enforcement Making the means explicit has a limit. No finite set of instructions can specify all situated work, and a longer instruction stack can conceal conflicts rather than resolve them. The architectural aim is to make the relevant purpose, constraints and sources available for competent reckoning. It must avoid equating that guidance with either complete understanding or mechanical prohibition. Instructions guide agent behaviour; they do not mechanically prevent every prohibited action. A resolver, sandbox, schema validator or operation engine can enforce particular predicates, subject to its entry points and coverage. Independent review supplies a different kind of examination. A human acceptance supplies an attributable commitment, not automatic error detection. A property-specific argument must identify what is prevented, by which mechanism and under which conditions; what is requested; what is detected afterward; and what remains dependent on practice. A conforming path-containment test establishes its tested boundary. Semantic correctness requires other evidence, and multiple layers can share the same mistaken assumption. Removing an obsolete validator can improve the system when its duty no longer serves the work. It does not remove the need to examine the actual permission, persistence and domain boundaries that remain. ### 9.2.5 LLM Capability Evolution Agent capabilities affect the appropriate delegation, instruction, recovery and verification arrangements. A procedure that once compensated for poor continuity may later interrupt useful work. Improved inference can make larger assignments practical; new capabilities can also expose new failure paths. A cheaper or degraded model may require more support or a narrower assignment. Benefit is not automatically monotonic in a fixed harness. A responsibility boundary can persist while the mechanisms supporting it change. Professional commitments, product properties and operational conventions each need examination on their own terms. Capability changes can warrant amendments to an assignment or a control; retaining a familiar mechanism is itself a design choice with consequences. Greater capability can warrant a revised assignment. It does not itself perform the person's judgment or grant institutional standing. ### 9.2.6 Over-Proceduralization and the Displacement of Attention Records, labels and procedures can help a practitioner attend to work; they can also become the objects of attention themselves. Polanyi's account of subsidiary and focal awareness helps describe this reversal [CITE:Polanyi1975]. The problem is a loss of integrated meaning when administrative particulars displace the engineering question they were intended to support. Decisions change their governing text; tools derive relationships; Git and PRs retain the committed change record. This arrangement is a design hypothesis: keeping current meaning accessible may improve continuity with fewer competing representations. Its value and costs require examination in actual work. The axiological assessment must include attention, delay, maintainability, user agency and the consequences of missed error. A control is valuable to the extent that its contribution to understanding, reliability or appropriate authority justifies those costs. Efficiency does not override every other value, but its costs cannot be excluded from the comparison. Some records remain essential: an identified result, a consequential human act or an uncertain operation outcome may need a durable trace. Machine execution history can also support diagnosis and recovery. Their function justifies them; routine duplication of the same account does not. No amount of recording can compel substantive judgment or guarantee that a person understands what they accept. --- ## 9.3 Generalizability ### 9.3.1 Across Engineering Domains Root's immediate subject is a development environment. The App's subject is agent participation inside an application, first through workflow authoring. SWBPIPE's subject is that participation in piping design and stress-model authoring. Keeping these questions distinct prevents success in one from being reported as proof of the others. The [App thesis](../../projects/chirality-app-v4/docs/thesis/README.md) can test whether discoverable operations, host authority and interaction support useful workflow work. The [SWBPIPE thesis](../../projects/chirality-piping/docs/thesis/README.md) can test whether domain meaning and numerical evidence remain faithful through human and agent entry points. Neither needs to reproduce Root's filesystem or all its methods to preserve the common orientation. Proliferation requires an economically useful transfer. Another application must adopt capabilities without rebuilding a large custom governance stack, and its users must understand the resulting operations. Shared schemas alone establish compatibility claims at most. Integration effort, preserved meaning, recovery and usefulness provide stronger evidence. The same kernel can orient all three theses while their architectural arguments remain practical and separately examinable. Appendix D supplies the concise conjecture; elaborating it further cannot substitute for evidence at a receiving application. The [shared introduction](INTRODUCTION.md#one-undertaking-three-perspectives) frames transfer as a relation among normative reasons, operative participation and domain evaluation. A failure in the receiving practice can therefore qualify the shared claim: an intelligible method may still impose excessive adaptation costs or omit a consequential domain distinction. Evaluating proliferation requires both the benefit transferred and the work needed to make it useful, rather than agreement in terminology alone. ### 9.3.2 Across Regulatory Jurisdictions The general distinction between producing information and taking responsibility for reliance can inform work in different jurisdictions. It does not transport the Alberta regulatory interpretation unchanged. Each application needs its own account of who may act, what review is due, which records matter and what constitutes authentication, if applicable. A successful translation may change controls substantially while preserving the framework's questions. Such a change is evidence of adaptation rather than necessarily a failure of the framework. ### 9.3.3 The Four-Pillar Framework as an Evaluation Lens The framework's value is discriminating attention. Ontology prevents a change in representation from passing unnoticed as a change in the work. Epistemology asks whether an available source supports the present claim. Praxeology asks whether local success reaches an intended use. Axiology asks whether that use and the means chosen remain worth pursuing. These relations help explain a failure that no single component's specification necessarily captures. At Root, the framework is useful when it exposes a missing distinction or an unsupported claim about how work is produced. An instruction has its own consumer, a dependency query its basis, a review its candidate and an acceptance its subject. These relations can be examined without inventing another administrative lifecycle: | Perspective | Question for an implementation | |---|---| | Ontology | Can participants distinguish the relevant work objects, relations and acts from their representations? | | Epistemology | Can they inspect the grounds, assumptions, conflicts and limits of material claims? | | Praxeology | Can authorized work proceed, be examined, corrected, integrated and recovered? | | Axiology | Are the purposes, tradeoffs and responsibilities explicit enough to assess the arrangement? | These questions constrain one another. A failure of recovery may reveal a missing distinction between an attempted and a completed action. A newly identified consequence may change the appropriate examination. A burdensome control may consume the attention needed for substantive review. There is no linear derivation from one pillar to a uniquely correct architecture, and no pillar's importance permits neglect of the others. The lens is non-exhaustive. It guides inquiry; it does not supply a complete ontology of work or guarantee that a system answering each question is adequate. ### 9.3.4 Epistemic Relations as Portable Concepts Claims, grounds, uncertainty, gaps, conflicts and decisions are useful concepts beyond this implementation. They need not all become fields in a register. Their representation should make the relevant relation discoverable where an undertaking needs it. Citation identifies a source; examination assesses something about its use; authentication is a particular attributable act. A cited claim can be wrong, a review can be limited, and a person can stand behind an identified result while disclosing uncertainty. These relations vary independently. What can transfer is a disciplined distinction among evidence, situated understanding and accountable reliance. The adequacy of a chosen representation remains a design hypothesis. It cannot externalize everything a knower brings to the work. ### 9.3.5 Capability-Invariance of the Authority Boundary **Greater agent capability does not itself transfer accountable judgment.** Write scopes, role arrangements and controls can change while that distinction holds. Smith's distinction between reckoning and judgment supplies the philosophical vocabulary [CITE:Smith2019]. Reckoning can be extensive and sophisticated: agents investigate, infer, plan, compare, criticize, propose, check and act within granted authority. This account does not equate reckoning with a simple calculation or require its execution through rigid procedures. Judgment is the person's situated and committed, world-involving engagement with what the work means, what matters and what they will stand behind. It shapes the question, relevant distinctions and grounds for reliance throughout an undertaking. A final approval is one possible expression of that engagement. Its record neither exhausts judgment nor guarantees its substance. The present professional allocation reserves accountable commitments to people. That allocation permits substantial agent initiative and continuing authority for ordinary work. It does not entail a read-only Agent 0, mandatory managerial tiers or renewed approval at every internal step. The agent responsibilities and the person's judgment concern different relations to the undertaking. Collins's account of socially acquired tacit knowledge contributes a further reason to examine context and participation rather than equating capability scores with accountable standing [CITE:Collins2010]. It does not settle the metaphysical possibilities of future artificial systems. Any proposed reallocation of professional responsibility would require an explicit institutional and normative argument; it is not an inference from improved output or delegated execution. Each property needs examination on its own terms. Improved models may change how responsibility is supported while leaving its current allocation in place. Conversely, a familiar mechanism may cease to support that allocation adequately. The framework makes these questions explicit and revisable. ### 9.3.6 Relationship to the AI Alignment Problem Model-level work seeks desirable behaviour through training, feedback and other interventions [CITE:Ouyang2022] [CITE:Bai2022]. Chirality examines the system of use: how outputs, operations, evidence and human acts are related in an undertaking. These concerns are complementary. The framework does not make a model truthful or safe under arbitrary deployment. Nor can an attribution record prevent every misuse of an output. The aim is to support appropriate action and reliance through useful distinctions, permissions, feedback and inspectable decisions. Whether these mechanisms succeed is an empirical question about the deployed system and its users. The reckoning/judgment distinction prevents the architecture from treating capability or an approval trace as a substitute for the person's engagement. It is a constraint on the account of the system, not a claim that an instruction can enforce all professional conduct. ### 9.3.7 Convergent Practice: Large-Project Delivery as Prior Art Root does not need a claim that these concerns are unique to agents. People have long managed dependencies, uncertainty and collective responsibility. Agents change the scale, pace and cost of producing candidates, and can make some intermediate structures themselves. That shifts which constraints dominate; it does not erase the need for interpretation and integration. The relevant novelty claim is therefore bounded by the implemented relation among these capabilities, not the invention of coordination or accountability. The structural views invite comparison with complex knowledge work, including large-project delivery. The correspondences below are a comparative hypothesis requiring independently sourced examination. They are not findings established by this case. The proposed correspondences can be stated without treating them as proof: | View | Question | Candidate comparison in project delivery | |---|---|---| | Tree | What is the work composed of? | Work breakdown and scope definition | | Dependency graph | What input or condition does this work need? | Precedence and interface planning | | Relationship network | What other material bears on this question? | Cross-references, sources and decision history | | Attention | How are contributions interpreted and brought into a coherent result? | Definition, coordination, review and synthesis | These are views and activities, not four required databases. A dependency graph need not be wholly acyclic: an executable ordering must respect actual prerequisites, while mutual design relationships can be legitimate cycles. A network of relevant material can be queried from current documents and history without an append register. Attention remains the integrating activity; its quality cannot be inferred from the presence of the other three representations. Neither agents nor people require every contextual fact to be written before work can occur. Hosts preserve different kinds of state, people carry tacit understanding, and participants can retrieve or investigate what is missing. The practical problem is recoverable access to the relevant basis, including uncertainty about that basis. Exhaustive institutional transcription is neither demonstrated as possible nor shown necessary. The comparison also must not equate a software acceptance control with an engineer's authentication, or presume that software work lacks consequential accountability. Analogies can identify design questions; their adequacy needs examination within the receiving practice. An independent comparison with project-delivery methods and a second practitioner's use would be more informative than further elaboration of the analogy by its author. --- ## 9.4 Future Work ### 9.4.1 Empirical Validation The Root question calls for assessment of comparable usable results, consequential escaped errors, recovery and total effort across the full production path. Time shifted from an agent to the owner, reviewer or next undertaking must not disappear from the account. Existing PR, CI and session evidence can support bounded observation without a new administrative register. Measures must account for work begun earlier, generated content, task mix and the owner's time. A controlled study with practitioners could compare arrangements on error detection, calibrated confidence and review effort. Self-reported confidence alone is insufficient: an interface can increase confidence without improving its grounds. The study should also examine whether participants understand the scope and limits of what they accept. ### 9.4.2 Enforcement at Consequential Boundaries Future engineering should target actual permission, operation, persistence and recovery boundaries. For example, an unknown action outcome calls for recovery semantics; an escaping write target calls for effective containment. Neither problem necessarily requires a general runtime governance engine. Each proposed control should state the consequential failure it addresses, its real enforcement path, its residual limits and its cost. Historical requirements for status propagation or universal receipts are not a standing implementation backlog. ### 9.4.3 Empirical Validation of Epistemic Representation A focused Root test can ask a fresh participant to locate the current condition for a real slice, inspect its supplier, identify the relevant consumers and continue within authorization. Observe whether the participant finds an actual source, relies on an obsolete summary, mistakes availability for suitability or initiates an unnecessary global search. This examines the value of deliverable sources and queries without treating metadata completeness as the outcome. A missing declaration can be corrected where useful; a sound inference may still require reading code or consulting the person. Counts of edges, citations or REVIEWED labels are not adequate proxies. Review has a parallel question: can the reviewer identify the candidate and the basis actually used, preserve valid evidence and detect a consequential shared premise error? The inquiry is whether the representation supports reckoning and human judgment in the actual undertaking. ### 9.4.4 Formal Analysis of Specific Properties Formal methods may help with narrowly defined properties such as path containment, state transitions, refusal of stale operations or duplicate application after interruption. Their value depends on a faithful model, verified implementation correspondence and coverage of the actual entry points. A proof about such a model cannot establish that an engineering claim is true, that the intended need is met, or that a person has exercised judgment. The scope of a proof must accompany its result. ### 9.4.5 Multi-Jurisdiction Regulatory Mapping A receiving jurisdiction's requirements should be investigated when a concrete deployment needs them. The result may require changed practice or product behaviour. This thesis neither pre-authorizes that use nor makes a broader regulatory survey a prerequisite for current development. ### 9.4.6 Multi-User Concurrent Execution Concurrent agents already contribute to repository work. The additional question is multiple accountable people acting on a shared undertaking: which authority applies, how concurrent edits are handled, and how each person understands another's commitments. Locks can address certain write conflicts. They cannot alone resolve conflicting purposes or responsibility. A useful test should examine a concrete shared activity and its recovery paths instead of assuming one coordination topology solves both mechanical and human problems. ### 9.4.7 Extension to Other Applications and Agent Platforms The strongest objection to proliferation is that each domain may require so much interpretation and custom integration that little of the shared framework reduces the work. That outcome would not make SWBPIPE useless; it would limit the broader architectural claim. The test is whether something valuable transfers while each application keeps its distinctive meanings, rather than whether both can use the same vocabulary. The App and SWBPIPE supply the first distinct receiving contexts. Root should be examined for what it enables them to reuse: current sources, selected methods, meaningful capability descriptions and an intelligible allocation of authority. The transfer need not include a common daemon, provider wrapper or record store. Native supplier capability can remain valuable precisely because Root avoids reimplementing it without a consumer. Successive application derivations add a longitudinal test. Does a new PRD-led implementation preserve relevant lessons, observed defects and useful behaviour, while removing obsolete topology? A new codebase or faster generation is not sufficient. Repeatedly rediscovering the same failure would qualify the claim of learning even if each rewrite satisfied its textual requirements. One human/agent domain operation can test common validation, meaningful refusal, inspectable outcomes and recovery. Broader proliferation should be inferred only as such paths accumulate in distinct applications, with their adaptation costs and remaining limitations visible. ### 9.4.8 A Second Accountable Professional, Outside Software The author remains a significant part of the functioning system. The owner sets direction, supplies domain interpretation and detects process burdens. A second accountable professional undertaking useful work without that continuous translation would test whether the framework transfers. The inquiry should observe what the practitioner can understand and use, which sources and tools help, what requires explanation, what consequential errors escape and what the work costs in attention and effort. A proposal, investigation or calculation package outside software would broaden the case. Success should not require the participant to reproduce the author's vocabulary or organization. --- ## 9.5 Summary Root supplies a coherent, revisable environment for human–agent production. Its strongest case is the connection between current artifacts, selective methods, queryable relationships, authority and continuing ownership through repairs. Its central danger is self-confirmation: producers and examiners can inherit the same framing, and a polished system of records can obscure the absent result. The case supports feasibility of specific production and correction paths. Productivity, sustained transfer and comparative reliability remain open empirical questions. The App and SWBPIPE examine different receiving contexts, each with its own subject and evidence. The useful trajectory is continued production with discriminating examination of the means. Root must make its own claims as accessible to correction as the claims in the work it helps generate.