Chapter 5 — Epistemic Architecture

---

5.1 Introduction

Three questions must remain separate:

1. What evidence or argument supports the claim?
2. What does that claim mean in the situation and for the intended purpose?
3. Who authorizes reliance on it, within what scope?

A citation can support a claim while leaving its relevance and intended use unsettled. Authentication records an act for which someone is accountable; it does not make the content true or establish that their engagement with it was adequate. Judgment shapes the work throughout, while agents extend reckoning through inquiry and action.

Knowledge resources are artifacts that support situated knowing: models, sources, calculations and returned findings help a participant inquire and act. How they are organized can preserve useful relationships without containing or reproducing everything a knower brings to the work.

---

5.2 The Problem: Absence of Intrinsic Epistemic Warrant

The literature documents model outputs that are fluent but unsupported or false [Ji2023] [Huang2023]. The important problem here is that fluent presentation does not establish grounds for reliance. Correct and incorrect claims can look similar on the surface, and neither a model's confidence nor a second model's agreement can by itself tell them apart.

This problem extends beyond generated prose. A cited source may be wrong, an equation may be applied outside its assumptions, a test may examine the wrong property, and a deterministic calculation may use the wrong units. Even faithful extraction establishes what a source says, not necessarily what holds in the world.

So engineering review needs access to the inputs, transformations, assumptions and checks relevant to the intended claim. Reviewers can target their effort when these are accessible, but neither labels nor provenance justify spot-checking a claim that matters without considering its failure modes. The amount and independence of examination depend on consequence, detectability and the evidence already available.

Chirality responds by making useful grounds and limits discoverable in the work's current artifacts and tools, without requiring a parallel description of every sentence. A calculation file with its inputs and an applicable oracle may be more informative than a prose register declaring that the calculation passed.

---

5.3 The Four Mechanisms

The four mechanisms below are design approaches to recurring problems. Their implementation must serve an actual reader, check or reliance decision. The identifiers cross-reference Appendix A, and Appendix E records the implementations as they evolve.

5.3.1 Mandatory Provenance (K-PROV-1)

The enduring requirement. An extracted claim that matters must be traceable to the source said to support it, or it must be clear that its support is missing. Provenance need not require a separate register.

Current application. Keep the source reference with the claim or its current technical basis. Identify a revision when changes in that source would matter. Tools can resolve references and compare revisions, but neither a successful resolution nor an unchanged hash establishes that the source is applicable or that the interpretation is correct.

The deliverable model keeps consumer needs and their conditions in deliverable.yaml, with supporting explanation in current source documents where useful. A query shows relationships from these declarations.

Worked example. An agent reads a scope document stating that detailed design requires approved structural analysis. The consumer's need can say:

Structural analysis results from DEL-01-01, suitable for the detailed-design basis. Source: scope document §4.2.3, identified revision.

If the agent also infers that foundation design needs geotechnical results, it can state that as an inferred need whose applicability remains to be resolved. It should not attach an unrelated scope citation and present the inference as an explicit requirement. The two needs differ in their grounds, even if both ultimately prove useful.

A reviewer can inspect the stated requirement, the proposed inference and the actual input. The presence of a supplier file or a merged PR does not establish that either need has been met.

5.3.2 No Invention (K-INVENT-1)

The enduring requirement. Do not present an invented value as an observed, extracted or accepted value. Make material unknowns apparent. This does not prohibit creative alternatives, exploratory calculations or explicit hypotheses.

A required missing input can block a particular calculation while leaving other work possible. The appropriate response is to identify that boundary, obtain or propose the missing input within authority, and continue unaffected work. A universal halt state for every uncertainty would obstruct inquiry.

Worked example. A vendor data sheet gives an operating temperature range of −20°C to +60°C and a flow rate of 500 m³/h, but no design pressure. A conforming extraction identifies design pressure as unknown. An agent may propose a candidate pressure for an explicitly bounded sensitivity study, but it must preserve that study's assumption and must not silently promote the candidate into the accepted design basis.

Instructions can require this behaviour, and checks can detect some violations, but neither instructions nor a TBD convention guarantee that every missing input has been recognized.

5.3.3 Conflict Surfacing (K-CONFLICT-1)

The enduring requirement. Do not conceal a material disagreement among sources or between evidence and the current basis. Show what conflicts, where it comes from and what action or decision it affects.

Suppose the scope requires operation at −30°C and a vendor sheet gives a −20°C limit. Averaging the values or choosing the more recent file does not resolve the conflict. The conflict's significance depends on whether the requirement, the equipment selection or the interpretation is wrong. An agent can investigate, propose alternatives and apply a resolution that is already authorized, but changing a reserved commitment or accepting risk belongs to the accountable person.

The resolution changes the current governing text or product, and its reason and material consequences belong in the work's existing change record. A conflict table can help with several related disputes, but a separate permanent ledger is not required for every discrepancy.

5.3.4 Epistemic Labeling

Labels can make an otherwise ambiguous distinction visible. They are annotations about a claim, not rankings of its truth or substitutes for its grounds.

An explicit source phrase, an assumption beside an equation, or a tool's unknown result can carry the distinction without an uppercase label. Use the form that helps the consumer interpret the work. Requiring a label on every non-trivial claim may create volume without making an assumption that matters easier to find.

Labels also differ from confidence estimates. A precise quotation can be labelled FACT while reporting a mistaken source, and a well-supported inference can remain an inference. Review must examine the grounds and purpose rather than treat labels as permission to rely.

---

5.4 The Epistemic Architecture as a Coherent Theory

The mechanisms address related questions:

1. Provenance: where do the claim and its asserted grounds come from?
2. No invention: what is actually missing, and what has been hypothesized?
3. Conflict surfacing: which statements or conditions cannot jointly govern?
4. Epistemic labeling: what kind of assertion is being made?

Their common aim is inspectability, which is constrained by ontology (which objects and claims are being examined), praxeology (how inspection leads to action or repair) and axiology (which failures and costs matter).

A useful architecture can expose a missing source, a failed oracle or an unresolved condition. But evidence that looks complete may share a mistaken premise, and a validator detects only what its rule examines. Smith's distinction between relative consistency of artifacts and correctness about the world remains relevant here [Smith1985].

So the theory yields a design question rather than a unique schema: what must be inspectable for the claims that matter in this work to be examined, and what representation makes that examination practicable?

---

5.5 Grounds, Examination and Accountable Reliance

Grounds, examination and reliance are linked relationships, but each can be revised independently:

• Grounds: a claim may have observations, calculations or arguments without a citation. A cited claim may have inadequate grounds.
• Examination: review has a subject, method, scope, revision and findings. A review can weaken a claim, reveal missing premises or become inapplicable after a relevant change.
• Reliance: acceptance or authentication links an accountable actor to identified content, scope and purpose. It is not a higher truth value.

These relationships can be represented where they matter, without assigning every claim a stored stage. A dependency query reports observable facts and unknowns. An agent assesses whether an input meets its condition within the assignment, and a person exercises judgment where accountable commitment is at issue. No aggregate count of citations or passed checks makes the work accepted.

The distinction between a descriptive record and a constitutive act matters. An agent's note that an owner approved something is a report that needs a source. Where an authorized process defines a person's action as acceptance, a host may capture that action and its scope faithfully. The person need not type every byte of the resulting record by hand. Conversely, a generated receipt or a hash match cannot invent the action or establish that the person's engagement was adequate.

Content identification supports the relationship. A fingerprint can show which bytes were named, and identity controls can support attribution, but they do not establish that those bytes were reviewed adequately, that the source actor was authorized, or that the conclusion was right. A later edit does not rewrite the historical act; it raises the question of whether the earlier acceptance applies to the changed work.

---

5.6 Comparison to Alternative Approaches

This comparison concerns what the cited approaches contribute. It does not claim that all implementations of them lack provenance or uncertainty reporting, nor is it a current benchmark ranking.

5.6.1 Reinforcement Learning from Human Feedback (RLHF)

RLHF trains behaviour using human preference signals [Ouyang2022]. Such training does not by itself provide the evidence needed for a particular engineering claim. Better model behaviour and better access to grounds can both improve work, but the examination still needed depends on the use.

5.6.2 Retrieval-Augmented Generation (RAG)

Retrieval makes source material available during generation [Lewis2020]. A retrieved passage can support a response, but its presence does not ensure faithful interpretation of the passage or reveal assumptions the response fills in beyond it. Citing specific sources and examining the resulting claim remain useful. Retrieval through files and tools can serve the same general purpose where the task requires it.

5.6.3 Post-Hoc Factuality Checking

Post-hoc checking can identify unsupported claims against a reference corpus [Min2023]. Its value depends on the corpus, the checking method and which claims it covers. A checker can itself be mistaken. Generation-time annotations and later examination are complementary options, not competing guarantees.

5.6.4 Constitutional AI

Principle-guided critique can improve behaviour [Bai2022]. Agreement with a principle does not supply empirical grounds for a domain claim, and a model's critique can share the original error. The useful distinction is between behavioural guidance and evidence appropriate to the claim, rather than an assumption that one makes the other unnecessary.

5.6.5 Summary of Positioning

---

5.7 Enabling Professional Reliance

The potential contribution is practical: a reviewer can inspect a calculation's basis, an input's revision, a material unresolved conflict and the claimed extent of a check, without having to depend solely on an agent's summary of those things. Existing product artifacts, source history and execution results may already provide the necessary evidence.

This support can reduce repeated investigation, but it cannot establish how much review a particular professional work product needs. For a consequential numerical claim, an independent calculation, oracle or method comparison may matter more than extensive annotation of claims. For an uncertain requirement, direct engagement with the owner may matter more than another test.

The empirical questions are whether material problems are found in time, whether the professional can understand the result and whether relevant actions can be reconstructed when needed. Passing a schema check is evidence about the schema check alone.

---

5.8 Summary

Chirality adopts a norm: claims should not exceed their available support, and accountable reliance must remain distinguishable from production and verification. It proposes inspectable sources, material unknowns, visible conflicts and useful annotations as ways to support that norm.

These are design hypotheses with costs and limits. They warrant records where a consumer, acceptance requirement or recovery need justifies them. Whether they are adequate depends on whether participants can examine the material grounds and act on findings in time for the intended use.

Agents can extend reckoning throughout the work. Human judgment shapes its purpose, interpretation, action and reliance, and it is not exhausted by an approval record. The architecture's contribution must be assessed through the work it enables and the failures it helps reveal, not inferred from how complete its own paperwork is.