# The Chirality Framework Chirality AI Ltd. — explanatory research, revised October 2026. **An account of professional accountability in agent-assisted work.** --- ## 1. Purpose Chirality supports purposeful work by making relevant objects and relationships intelligible, the grounds and limits of claims inspectable, action and correction practicable, and consequential commitments attributable. The mechanisms that provide these conditions remain open to revision through use. Professional accountability concerns what may be relied on, by whom, on what grounds and under what commitment. Three kinds of statement must remain distinct: - **Normative commitments** identify values and responsibility boundaries the undertaking chooses or is obliged to preserve. - **Design hypotheses** propose mechanisms that might support those commitments. - **Empirical findings** report what observation establishes under identified conditions, including failures, costs and limits of transfer. A coherent design is not evidence that it works. A successful check supports its tested claim, not every claim about the design. --- ## 2. The Nature of Knowledge in Professional Practice ### 2.1 Knowledge Belongs to a Knower Information can be recorded, transmitted, copied, cited, compared and identified without being the knowledge of any particular person. Knowledge, in this framework's usage, is a situated achievement of a knower. The framework does not use *knowledge* as a guarantee of truth. A person may know wrongly, incompletely or provisionally, and may revise what they know. This differs from accounts that reserve the word for true knowledge. The practical argument can also be read in terms of situated understanding: evidence and review discipline a person's understanding without transferring it into an artifact. `Claim` and `Warrant` identify an assertion and its grounds. They make a potential basis for knowing inspectable. Neither is a knower, and their conjunction does not define knowledge. ### 2.2 Configurational Multiplicity Identical information may occasion different knowledge for different people, or for one person in another context, purpose or time. Knowing is situated; its variation need not imply a change in the information. Chirality calls this **configurational multiplicity**. Schemas, taxonomies, semantic lenses and graphs can preserve distinctions and direct inquiry. They do not exhaust what a person may perceive in the work. This openness does not make interpretations equally well supported or relieve a person of answering to evidence. ### 2.3 Authentication and Accountable Reliance Professional authentication is an attributable act concerning identified information. The accountable person examines a defined work product and accepts responsibility for reliance within an identified scope and purpose. The subject must be identifiable. A specific content revision or hash can establish identity; the applicable professional and product requirements determine the binding. Authentication confers **accountable-reliance status** within that relation. It does not create knowledge, make the information true or determine what another person must know from it. A record can evidence the act. The presence of a signature, approval field or hash does not establish that sufficient examination or committed engagement occurred. Judgment also shapes the question, intended use, relevant distinctions and choice of evidence before acceptance. It is not exhausted by a final approval. ### 2.4 The Chirality of Knowledge The framework's primary use of *chirality* names the **accountability gap** between externalizable information and accountable knowing. An artifact, provenance chain or approval record is not identical with the person's knowing. The metaphor marks this non-identity; it supplies neither a geometry of knowledge nor a new chiral structure for every architectural distinction. The gap is workable. Information, grounds, uncertainty and attributable acts can support a person's engagement without replacing it. Another citation or a more capable model does not, by itself, perform that person's commitment. This gap differs from the operational `Gap`: a missing warrant or unresolved input that further inquiry can address. The accountability gap is not a missing field to be completed. ### 2.5 Agency, Delegation, and Accountability An artificial agent combines a model, instructions, context and tools through a host. It can investigate, infer, compare, propose, assess, plan, check and act within authority. These activities can be sophisticated, adaptive and sustained. The framework describes these capabilities as **reckoning**, following Smith's distinction from **judgment**. Judgment is the person's situated, committed, world-involving engagement with what the work means, what matters and what they will stand behind. It operates throughout purpose, framing and action, not only at approval. The allocation of professional responsibility concerns a different question from capability. An agent may perform extensive authorized work without occupying the licensed professional's role. In this framework, it does not authenticate a work product or assume duty of care. This states the allocation in the present professional setting; it is not a proof about the possible cognition or institutional standing of every future machine. Greater capability can change appropriate delegation, supervision and verification. It does not itself transfer accountable judgment. Conversely, reserving judgment does not require Agent 0 to be read-only or every ordinary design refinement to receive a new approval. ### 2.6 Teams, Project Management, and Governance Work requires continuity of purpose, ownership of integration and a way to coordinate dependencies. It need not require several agents. Roles describe contributions; responsibilities do not prescribe a staffing chart. In the current repository, the agent working with the owner holds Agent 0's responsibilities. Managers and bounded executors are engaged when useful. TASK does not delegate. An owner of work continues through ordinary repairs, within the assignment and the host's actual permissions. Current commitments, design and dependency conditions live with the deliverables. Tools derive relevant relationships and observed facts. The conversation supplies present intent; Git and PRs retain changes and reasons. A recorded dependency does not prove readiness, and a missing edge does not prove independence. Additional records are useful only where review, use, acceptance or recovery needs them. The normative, operative and evaluative perspectives concern one undertaking. Commitments orient action; operations make action possible; evaluation examines what occurred and what it supports. Their relation is iterative: a finding can revise an operation, a method or an adopted criterion through the appropriate authority. These are overlapping perspectives, not a prescribed hierarchy of agents or a fixed sequence of approval stops. The [shared introduction](INTRODUCTION.md) develops this relation across the three theses. ### 2.7 What the Framework Does and Does Not Cover Professional reliance involves agency, organization and authority as well as information. A person may need to know who acted, what changed, what evidence was examined and which limitations remain. The framework addresses these relationships. It neither claims a complete theory of human life nor establishes that one implementation is sufficient for regulated practice. Legal and professional requirements apply in their own jurisdictions and contexts. Product contracts and actual examination must establish the conditions for a particular use. ### 2.8 Normalization Table | Element | Kind | Contribution or standing | Limit | |---|---|---|---| | AI agent | Artificial operative participant | Reckoning and action within authority | Does not acquire professional responsibility from capability | | Accountable professional | Person under applicable professional obligations | Judgment, supervision and attributable reliance | Approval does not guarantee correctness | | Claim | Assertion about a subject | Makes content available for examination | Plausibility is not a warrant | | Warrant | Grounds relevant to a claim | Supports assessment within its actual scope | Citation presence does not establish relevance or adequacy | | Ruling | Attributable decision | Settles a choice within the issuer's authority | Does not manufacture evidence or resolve a factual conflict by decree | | Work product | Artifact used for a purpose | Contains or points to relevant content and grounds | Completion and technical validation do not establish acceptance | --- ## 3. What the Architecture Must Provide ### 3.1 The Generative Principle A person needs access to the material grounds and limits of the reliance they are asked to accept. Architecture can make these easier to find, examine and relate to the actual work. That is a design obligation within this framework, not a guarantee that the person will understand or examine them adequately. Inspectability concerns useful access. It does not require a separate record of every activity. A current artifact, a test result or an existing change record may already supply what is needed. Repeated representations can obscure consequential decisions and displace attention from the work. ### 3.2 The Four Pillars | Pillar | Enduring question | What it can expose | |---|---|---| | **Ontology** | What kind of thing or relationship is this? | Confusion between a proposal, an operation, a result and acceptance | | **Epistemology** | What supports the claim, and what are its limits? | Unsupported inference, stale evidence or inadequate examination | | **Praxeology** | How does intent become usable work through action and correction? | Unowned integration, blocked independent work or uncertain outcomes | | **Axiology** | What matters, to whom, and with what consequences? | A narrow optimization that neglects safety, agency or useful attention | The questions are mutually corrective. A failed recovery attempt can reveal an omitted outcome in the ontology. A newly identified consequence can require different verification. Experience can show that a supposed requirement was an implementation habit. The four questions need no fixed document set, review sequence or permanent process topology. Control costs belong inside the evaluation. Attention, delay, maintainability and user agency matter alongside reliability. A control earns its place when its contribution justifies those costs. Applicable obligations and adopted constraints govern what can be accepted; negotiable preferences support choices among permissible alternatives. Uncertain premises require examination rather than a favourable ranking. A local improvement can impose costs on construction, operation, maintenance or another person's work. No universal score settles these relationships. Judgment concerns which consequences matter in the actual undertaking and which mechanisms serve its obligations. ### 3.3 The Epistemic Architecture Epistemology receives particular attention because plausible outputs can conceal weak grounds. It does not subordinate the other pillars. Purpose determines which uncertainty matters; representation determines what can be examined; practice determines whether examination and correction can affect the result. #### 3.3.1 The Epistemic Ontology | Concept | Meaning | |---|---| | **Claim** | An assertion about a subject | | **Warrant** | Relevant grounds, such as a source, observation, calculation, test or argument; model confidence alone is not a warrant | | **Status** | An explicitly scoped classification useful to examination; not a substitute for the grounds | | **Gap** | Missing grounds or information, with its effect on the current work made visible | | **Conflict** | Incompatible claims whose scope, sources or meanings require examination | | **Ruling** | A decision by the relevant authority where a choice is needed | FACT, ASSUMPTION, PROPOSAL and TBD remain useful labels in contexts that adopt them. They are not an exhaustive account of knowledge or a universal requirement to annotate prose. An observation requires context; an assumption remains provisional; a proposal is not an adopted commitment; TBD identifies missing information. A human ruling can select a course of action under uncertainty without turning the uncertain premise into a fact. #### 3.3.2 Mechanisms and Their Enforcement Limits Provenance, explicit uncertainty, conflict surfacing and proportional examination are useful mechanisms. Their implementations vary by consumer and consequence. Some are instructions; others are schema or runtime checks. A check for a source field establishes that a field exists, not that the source supports the claim. A prohibition on invention does not ensure that fabrication never occurs. #### 3.3.3 Evidence and Acceptance Evidence may be strengthened, contradicted or become inapplicable. Review has a subject, method and scope. Acceptance is an authorized act concerning an identified subject and purpose; professional authentication has additional obligations in its applicable setting. These relations do not form a universal progression. Their historical combination in a warrant lifecycle is discussed in [Appendix E](appendix_e_development_history.md). The practical questions are: what grounds support this claim; what examination has addressed it; and what reliance, if any, has a person accepted? No general deliverable status is required to answer them. Authentication does not automatically release a product, approve later changes or broaden its scope of use. ### 3.4 The Fractal Property The same questions can recur at several scales: one numerical result, a work product, an application or the agent-assisted undertaking itself. Their recurrence calls for attention to the particular subject, not a prescribed set of forms. Recurrence is valuable when it reveals a consequential distinction at the scale being examined. It is not proof of completeness or a reason to reproduce the same forms throughout a repository. ### 3.5 The Generative Relationship to Systems Engineering Systems engineering supplies candidate methods for preserving these distinctions: configuration management can identify a subject; verification and validation can test specified claims; interface design can constrain operations; human factors can assess how a person interprets and controls the work. The pillars guide the selection and evaluation of these methods. They do not uniquely derive them. Each mechanism requires a reason tied to the undertaking, then evidence of its adequacy and cost. Changes in model capability can make some supervision mechanisms unnecessary while leaving the underlying responsibility unchanged. --- ## 4. What the Architecture Cannot Do ### 4.1 The Boundary Architecture can organize information, perform operations and preserve evidence. It cannot perform a particular person's knowing or commitment on their behalf. Neither a complete form nor a passing check establishes adequate professional engagement. The person still interprets what matters and stands behind the reliance they accept. ### 4.2 The Human Responsibility An agent can extend reckoning across investigation, design, implementation, examination and repair. The current professional allocation retains judgment and duty of care with the accountable person. This boundary grants no new permission and imposes no fixed agent hierarchy. Systems must not attribute an approval that the person did not give. Where an approval is required, its scope and subject must remain clear. The retained product boundary K-AUTH-1 applies to its adopted consumers; this framework explains the reason without changing that contract. --- ## 5. The Central Distinction Useful output and accountable reliance are related but different achievements. A productive tool can improve quality and reduce effort. An accountable work arrangement also enables a person to examine what matters and identify what they will stand behind. These aims can support each other. Removing redundant records can improve both throughput and examination. Conversely, faster output can increase review burden or conceal uncertainty. Neither administrative volume nor production speed alone establishes the quality of the arrangement. --- ## 6. The Axiom *To know what you know, why you know it, and why you know why.* The first clause directs attention to the subject of knowing. The second asks for its grounds. The third makes the grounds and commitments of the framework itself open to examination. It invites reflexive inquiry rather than an infinite chain of required records. Chirality must therefore distinguish its own commitments, hypotheses and findings. Its arrangements must answer to evidence about usable results, failures, recovery and total effort. A coherent alternative or successful local correction cannot establish superiority in every application or professional setting. --- ## 7. The Closing Principle AI can extend professional reckoning without inheriting professional judgment or responsibility merely by becoming more capable. People can grant extensive authority for action while retaining consequential commitments. The framework's coherence lies in preserving relevant distinctions and responsibilities while examining how best to support them. Its implementations remain contestable. Their adequacy must be assessed through actual work, including the costs they impose and the failures they permit.