Purpose and Interface Scope

The complete evaluation results for Chapter 5 (the standardisation schema) are documented here, and the supplementary data package deposited with this thesis is described in the final section. The evaluation addressed three substantive questions: whether the schema resolves clause-level ambiguity in the NDIS SDA Design Standard 2019, whether resolution concentrates differentially across clause classes in a pattern consistent with the schema’s design intent, and whether the two-channel architecture is warranted by measurable complementarity between the text and figures channels. All reported metrics are grounded in artefacts held in the data package described in the final section; evidence registers mapping reported results to specific package files are listed there.

Text-Channel Evaluation Corpus and Integrity

The text-channel evaluation corpus is defined by 611 serialised clauses drawn from the NDIS SDA Design Standard 2019, covering 25 design elements. Clauses were serialised using the schema documented in Appendix: Standards Serialisation Schema and assigned clause-class labels (design_requirement, rationale, applicable_to, other). Corpus integrity baselines (structural completeness checks, identity-stability analysis, and cross-layer coverage measures) are reported in Appendix: SDA Corpus Integrity Metrics.

Text-Channel Ambiguity Resolution Results

The central evaluation measure is the per-clause ambiguity-delta: the reduction in measurable ambiguity achieved by the schema relative to an unstructured baseline. Across the 611-clause corpus, the mean per-clause ambiguity-delta is 0.2514. This is the aggregate resolution rate against which the pre-registered success threshold of 0.25 (Section 4.5) is applied: it clears the threshold, but by a thin margin of 0.0014, so the pass is reported here as narrow rather than comfortable. The aggregate full-schema delta against a null-model baseline (no schema applied) is 0.375, compared with 0.007 for the null model. The minimal-schema variant yields 0.307, establishing that clause-class annotation contributes +0.301 of the total gain and deontic annotation contributes an additional +0.067.

Table A5.1: Ambiguity-delta by schema variant

Resolution is concentrated in design_requirement clauses, the class carrying direct normative force, with an ambiguity-delta of 0.570. The rationale class achieves 0.475; applicable_to achieves 0.023; other achieves 0.000 by design, as the schema excludes residual clauses from resolution logic. This distribution confirms that the schema resolves ambiguity where the consequences of ambiguity are highest, rather than across the corpus indiscriminately. In summary, the clause-class-differentiated resolution profile is the principal evidence that the schema’s design intent, to concentrate resolution power where normative stakes are highest, is realised in practice. The next section reports the deontic force analysis that extends this picture to the modal dimension of the corpus.

Table A5.2: Ambiguity-delta by clause class

Structural resolution and deontic resolution rates are 90.6% and 90.9% respectively across the 611-clause corpus, indicating that clause-role and modality profile are preserved through the schema in the large majority of cases.

Deontic Force Analysis

The corpus is defined by 164 modal clauses, and deontic outcomes require three-way decomposition. Of the 75 coverage-gap cases, ontology coverage is insufficient to assess force; these cases represent a transparent design boundary rather than force loss. Of the remaining cases, 61 represent actual force collapse and 28 represent force-preserved resolution. This decomposition separates ontology under-coverage from genuine modality loss: a distinction that a single aggregate collapse rate obscures.

Lexical ambiguity over the corpus is a census property, reported here as a plain proportion rather than estimated from a sample. Of the 204 lemmas that meet the frequency threshold of at least three appearances, 121 (59.3 per cent) are polysemous as the corpus stands. The polysemous set includes built-environment fundamentals (door, wall, height, basin) and spatial-category terms such as space and area; because these are also among the most frequent lemmas, their polysemy is the single largest source of the interpretive burden the schema must address explicitly. The burden concentrates where it matters most for compliance: of the 54 fallback terms the schema retains, 107 of their appearances fall within the 164 modal clauses. Full per-term polysemy counts are in Appendix: Polysemy Metrics and Confidence. The text-channel ambiguity and deontic results together show the schema reducing interpretive uncertainty precisely where normative stakes are highest while leaving the residual burden visible rather than hidden. The figures-channel results in the next section, and the cross-channel complementarity they reveal, are what warrant the two-channel architecture.

Figures-Channel Results and Cross-Channel Complementarity

The figures channel provides 189 design requirements extracted from SDA Design Standard figures, complementing the 611-clause text corpus. Of the 189 figure-derived requirements, 38.6% (73 of 189) have no equivalent in the text channel. Conversely, 86.4% of text-channel design requirements have no figure equivalent. The figures-channel ambiguity-delta is 0.9148, substantially higher than the text-channel mean of 0.2514. Together, these figures establish that the text and figures channels carry largely disjoint normative content and differ substantially in their amenability to structured resolution, warranting the two-channel architecture rather than single-channel extraction from either source alone. Overall, the figures-channel results and cross-channel complementarity data constitute the central evidentiary basis for the schema architecture decision documented in Chapter 5. The next section presents the robustness and validation findings that confirm the stability of these results across corpus partitions.

Stability and Sensitivity Checks

Because the corpus is a complete census rather than a probability sample, stability is reported as descriptive structure across partitions of the corpus, with no interval estimate or significance test attached. When the 611 clauses are divided into ten equal partitions, on average 94.96 per cent of the terms in any one partition also occur in the others, so the schema vocabulary is shared across the corpus rather than confined to any single region of the standard. Structural coupling between primitives is reported descriptively rather than tested against a null: the co-occurrence counts for every primitive pair are deposited in the data package, and they show coupling concentrated among a minority of pairs rather than spread evenly. The one design parameter that could have driven the results, the minimum-frequency threshold, is shown to be non-load-bearing by reporting the analysis at each adjacent cut-off from one to ten appearances: no finding in this appendix changes across that range. The next section situates these results against established building and housing ontologies.

External Benchmarking

Comparison against established building and housing ontologies establishes the superiority of the schema-first approach. The SDA Design Standard serialisation achieves 100.0% clause coverage of the corpus by construction. Applied to the same clause set, IFC4 achieves 32.1% coverage and BOT achieves 20.8%, leaving approximately two-thirds of normative content unrepresented under an ontology-extension strategy. The systematic literature review underpinning Chapter 5 identified 15 of 134 floor-plan representation papers as addressing accessibility concepts, confirming that accessibility representation constitutes an under-served area in the field. Therefore, the external benchmarking evidence confirms that the schema-first approach achieves normative coverage that established ontology-extension strategies cannot match for this instrument, and this establishes the comparative justification for the design decision to develop a purpose-built serialisation schema rather than extending an existing ontology.

Supplementary Data Package

The Chapter 5 data package is deposited with this thesis in two forms: a single compressed archive and an expanded directory for direct file access. Its deposit locations are recorded in the Data Package Manifest below; an examiner can locate every component of the package from that manifest. All file references elsewhere in this appendix are given relative to the root of the deposited package (the ch5-artefact-bundle entry in the manifest).

Data Package Manifest (package-relative)

The interactive explorer, opened in any web browser, provides entity-browsing, triple inspection, and design-category filtering across the full serialised dataset without requiring a server. Within the expanded package, Table A5.3 names the canonical datasets and their record scope.

Table A5.3: Canonical datasets in the Chapter 5 artefact bundle

Four flat-file spreadsheet exports accompany the package for tabular inspection: entities, triples, polysemy, and requirements-by-category. Semantic exports for graph-tool and semantic-web reuse are provided as a directed knowledge graph in GraphML and as an RDF Turtle ontology (planimetric.org namespace).

Field-level documentation for every field in every dataset is provided in the package’s data dictionary. Regeneration programs accompany the package and can be run against the canonical datasets to reproduce the derived flat-file and semantic exports.

Provenance and Versioning

Evidence Appendices

The detailed findings for each evaluation dimension are presented in dedicated examiner-facing appendices:

• Appendix: Chapter 5 Figures-Channel Ambiguity Analysis: five-dimensional ambiguity model and per-dimension resolution rates for the 406-entry figures corpus
• Appendix: Chapter 5 Polysemy Burden Assessment: WordNet lexical polysemy, dimensional polysemy, and categorical absence polysemy across the 56-entity vocabulary
• Appendix: Chapter 5 Cross-Channel Validation Results: match type distribution and directional asymmetry between text and figures channels
• Appendix: Chapter 5 Predicate Coverage and Deontic Force: deontic force distribution, modal term inventory, force by field type and SDA category, and predicate vocabulary
• Appendix: Chapter 5 Data Pipeline and Reproducibility: pipeline architecture, file inventory, verification procedures, and reproducibility constraints
• Appendix: Chapter 5 SDA Corpus Integrity Metrics: structural completeness, identity stability, and cross-layer coverage for the text-channel corpus

Raw evidence files mapping reported results to specific data package artefacts are also deposited in the package’s evidence/ subdirectory (its location within the deposited thesis is recorded in the Data Package Manifest above), for reference by researchers accessing the data package directly.

Appendix: Chapter 5 Reference Tables

These tables are referenced from Chapter 5 (the standardisation schema). They were relocated from the chapter main text to keep the in-text body within the length band while preserving the full data (no content cut). All figures are census quantities over the serialised SDA Design Standard, reported descriptively with explicit denominators.

Analytical methods (Section 5.4)

The six analytical methods, each targeting a distinct dimension of ambiguity.

Baseline dimension summary (Section 5.11)

The six empirical-baseline dimensions in compact form.

Modal-operator frequency distribution (Section 5.14)

Six modal operators across four force categories; 164 modal clauses of 611 total (modal rate 0.2684).

Polysemy confidence tiers (Section 5.14)

Cross-method agreement assigns confidence tiers over the 204 eligible lemmas.

Design Science Research iteration register (Section 5.8)

Nine documented DSR iterations (0-7b) across four analytical cycles; the design-search history behind the standardisation schema (visualised in the Design Science Research iteration history for the standardisation schema figure).

Chapter 5 SDA Corpus Integrity Metrics

This appendix documents the corpus integrity baseline for the text-channel serialisation corpus evaluated in Chapter 5. Corpus integrity refers to the set of structural completeness, identity stability, and cross-layer coverage properties that must be established before ambiguity analysis results can be interpreted with confidence. If the corpus were structurally incomplete, contained truncation artefacts, or exhibited systematic identity instability, the resolution-rate figures reported in Appendix: Chapter 5 Evaluation Results and Data Package would be confounded by data quality defects rather than reflecting genuine schema performance. These metrics confirm that the corpus meets the minimum quality thresholds required for the analysis to be interpretable.

Structural Completeness

The text-channel serialisation corpus comprises 611 total records. Structural completeness checks confirm that zero records are missing required keys, zero records contain truncation markers, and zero records have malformed clause headers. These three checks address the three most common failure modes of large-scale document extraction: key omission (where a required field is absent from the serialised record), truncation (where the source text was cut short during extraction), and header malformation (where the clause identification scheme is inconsistent). Clean results across all three checks establish that the 611-record corpus is structurally sound and that no data quality defects confound the ambiguity analysis. In summary, the structural completeness baseline confirms that the corpus is fit for the ambiguity analysis that follows. The next section examines whether record identity is stable across the corpus: a property that structural completeness alone does not guarantee.

Identity Stability

Identity stability analysis examines whether the same serial identifier is associated with different textual content across the corpus: a signature of referential instability where the same identifier is used for distinct requirements. The analysis detects 8 duplicate IDs that reference different text content, and 0 duplicate IDs referencing identical text content. The presence of 8 identity-unstable records justifies the record-level identity handling implemented in the serialisation schema: the schema assigns a stable internal identifier to each requirement based on its content rather than relying solely on the source document’s nominal clause numbering. The 0 identical-text duplicates confirm that the instability is semantically meaningful (distinct requirements sharing an identifier) rather than a simple copy-paste artefact.

The 8 unstable identifiers affect 8 of 611 records (1.3% of the corpus). This rate is low enough that the corpus supports aggregate ambiguity analysis, but high enough to warrant the identity-handling mechanism as a design requirement rather than an optional feature. Building on this identity stability finding, the next section reports cross-layer coverage, the proportion of serialised clauses that can be linked to the ontological layer, which reveals the deliberate design boundary between the text corpus and the spatial-dimensional ontology built for figure-based triples.

Cross-Layer Coverage

Cross-layer coverage measures the proportion of serialised text requirements that can be linked to entries in the ontological layer: the layer that provides semantic type information, entity classification, and relational context for downstream processing. Of the 603 unique serial IDs in the corpus (derived from the 611 total by removing the 8 identity-unstable duplicates), 30 are linked to ontology entries, yielding an ontology coverage ratio of 0.0498 (4.98%).

The coverage ratio of 4.98 per cent reflects a deliberate design constraint. The ontological layer was developed to support the figures-channel triple extraction pipeline, and its scope therefore excludes the broader 611-clause text corpus. The text corpus covers a broader regulatory scope (procedural requirements, conditional applicability rules, and external standard references) that does not map readily to the spatial-dimensional ontology built for figure-based triples. The 4.98% coverage ratio evidences representational under-coverage and is declared as a design constraint in Chapter 5 rather than treated as noise.

Modal Carry-Over

Modal carry-over measures the proportion of text-corpus clauses identified as modal (carrying explicit deontic force) that are also covered by the ontological layer. The corpus contains 164 serial modal clauses. Of these, 9 are covered by the ontological layer, yielding a modal coverage ratio of 0.0549 (5.49%).

This ratio is slightly higher than the general ontology coverage ratio (5.49% versus 4.98%), indicating a marginal concentration of ontological coverage in modal content. The finding is consistent with the ontological layer’s design intent: entries were prioritised on the basis of spatial and deontic salience. However, the absolute coverage remains low, confirming that the text-channel modal analysis must be conducted primarily through direct linguistic analysis of the serialised clauses rather than through ontological inference. Overall, the modal carry-over finding reinforces that the text and ontological layers are intentionally scoped to complementary purposes, not to comprehensive mutual coverage.

Table A5-CI.1: Corpus integrity summary

Source: structured analysis of the Serialised Text Requirements Corpus and the associated ontology-linkage tables (see Data availability above).

Four corpus counts recur across the Chapter 5 appendices on different bases, and Table A5-CI.2 reconciles them so no count is read against the wrong denominator.

Table A5-CI.2: Corpus count reconciliation

The 611 → 603 transform and the 164 decomposition are reproduced above from App G’s own figures. Three derived quantities are registered to the sibling data-quality reconciliation rather than adjudicated here: the exclusion rule that maps 611/603 to the 605 trigram/POS base; the ontology-coverage numerator (30 linked IDs); and the modal-coverage numerator (9 ontology-covered modal clauses). The ratios above are stated consistently with their printed numerators and denominators.

The corpus integrity checks collectively confirm three properties necessary for the Chapter 5 evaluation: (1) the corpus is structurally complete and free of extraction artefacts; (2) referential identity instability is present and explicitly handled by the schema design; and (3) cross-layer coverage asymmetry is measurable and declared as a design constraint rather than hidden as noise. Taken together, these three integrity properties establish the evidential foundation on which the resolution-rate and deontic-force analyses in the companion appendices rest. This establishes the corpus as a sound and interpretable basis for the evaluation claims advanced in Chapter 5.

Trigram and Part-of-Speech Structural Metrics

Trigram and part-of-speech structural analysis is documented here for the standards representation corpus in Chapter 5. The analysis was performed on 605 processed clauses to characterise the phrase-level structural patterns that differentiate clause classes and establish the structural regularity needed for schema design. This 605-clause base sits between the 611 total records and the 603 unique serial IDs; the exclusion rule that maps the corpus to it is registered to the data-quality reconciliation (see Table A5-CI.2 in Appendix: Corpus Integrity Metrics) and is not reproduced here. Both aggregate counts and per-class dominant patterns are reported.

Six aggregate metrics are reported for the structural analysis corpus. | Metric | Value | | — | — | | Processed clauses | 605 | | Distinct trigrams extracted | 4,556 | | Retained trigrams (minimum frequency threshold) | 568 | | Dominant applicability-frame trigram | clause applicable to | | Dominant requirement-frame trigram | clause design requirement | | Dominant rationale-frame trigram | clause rationale for |

568 trigrams are retained from 4,556 distinct trigrams (approximately 12.5 per cent). This ratio reflects the high phrase-level variation characteristic of regulatory documents, where clause wording is rarely standardised below the clause-class level. The retained trigrams are recurrent phrase structures that appear across multiple clauses and carry interpretable structural signals. The dominant trigrams listed above identify the lexical anchors through which the three primary clause classes are consistently marked in the corpus. Overall, the trigram profile confirms that clause-class recovery is feasible for the three primary frames. The residual other class lacks a dominant structural marker and cannot be recovered from phrase structure alone. The next section documents the clause-class structural roles that the trigram analysis informs.

Clause-Class Structural Profile

Four clause classes are identified by the structural analysis. Each class carries a distinct structural role in interpretation, and conflating classes produces inferential errors in downstream schema assignment.

Interpretation

Three interpretive conclusions are drawn from the structural metrics in Chapter 5. First, clause framing performs semantic work that cannot be ignored in representation design. Dominant trigrams demonstrate that structural markers are systematically associated with clause class. Clause class is therefore partially recoverable from phrase structure and must be explicitly assigned rather than assumed uniform. Second, structural ambiguity is pattern-level as well as term-level. The high count of distinct trigrams (4,556) relative to the retained set (568) shows that phrase-level variation is substantial. Ambiguity arises from structural sources as well as lexical ones. Third, clause class is mandatory schema metadata for avoiding inferential conflation. Without explicit class assignment, design requirements, rationale statements, and applicability conditions carry equal weight in downstream processing. Representation errors accumulate through the artefact chain as a result. Taken together, these three conclusions establish that the clause-class annotation layer is a substantive design requirement. It is grounded in measurable structural properties of the corpus. The trigram and part-of-speech evidence presented here provides the structural justification for this field. The standards serialisation schema mandates clause-class as required metadata for every governed row.

Foundational Primitive Mapping Coverage

This appendix documents the foundational primitive mapping coverage achieved in Chapter 5 of the thesis. The mapping operation assigns each high-frequency term in the standards representation corpus to one of fifteen foundational spatial-semantic primitives. Complete coverage at the selected frequency threshold was achieved, but the distribution across primitives reveals a pattern of residual conceptual burden that must be interpreted alongside the aggregate coverage figure.

The coverage ratio of 1.0000 confirms that every term at or above the selected frequency threshold received a primitive assignment. This result is an artefact of the foundational abstraction discipline: the primitive set was designed to be collectively exhaustive at the level of spatial semantics, and element was designated as the catch-all primitive for built-component terms that could not be assigned to a more specific category with adequate confidence. The fallback-to-element count of 54 directly expresses this residual burden. Of the 204 mapped terms, 54 (approximately one quarter) were assigned to element not because their semantic content was unambiguous but because no more specific primitive could be justified under the confidence requirements of the schema. High coverage and meaningful residual uncertainty therefore coexist in this dataset. In summary, the mapping coverage profile confirms that the schema achieves complete assignment at the selected threshold while concentrating adjudication pressure in the element primitive, and this establishes the basis for the dual-reporting norm that Chapter 5 adopts throughout its evaluation.

Primitive Inventory

The fifteen candidate terms considered in the mapping are listed in the Standards Serialisation Schema Specification appendix. Following ablation analysis in Chapter 5, the zone term was removed from the operative set, on the cognitive account a composite (space bounded by boundary) whose work the space primitive and the bounded_by operator already carry, reducing the inventory to the fourteen entity terms (seven primitives and seven composites) used in the final schema. The ablation result is documented in the schema specification appendix, where information-loss scores justify each retained and removed term.

Interpretation

Three interpretive principles govern the use of this coverage profile across the thesis. First, coverage alone is not a sufficient semantic quality claim. A coverage ratio of 1.0000 is expected given the design of the foundational primitive set and does not constitute evidence that every assignment is semantically precise. Second, residual load must be reported alongside coverage. The fallback-to-element count of 54 is the critical figure for understanding where adjudication pressure accumulates in the corpus and where schema governance is most exposed to approximation. Third, primitive assignment must preserve ambiguity visibility. Where term-level confidence is limited, the schema contract requires that ambiguity indicators be retained at the row level rather than suppressed by a nominal coverage claim.

These principles underpin the dual-reporting norm adopted in Chapter 5: coverage and residual burden are always stated together, and the combination is interpreted as evidence about schema governance quality rather than as a simple accuracy score. Overall, these three interpretive principles apply to every coverage claim in the thesis: completeness figures are always read alongside the fallback count, ambiguity visibility is maintained at row level, and no coverage claim is advanced as an accuracy claim without declaring the residual burden that accompanies it. Therefore, the mapping coverage data documented here functions as an integrity baseline for interpreting the ambiguity-delta and primitive-assignment results in the companion evaluation appendices.

Chapter 5 Predicate Coverage and Deontic Force

This appendix reports the deontic force classification and predicate vocabulary analysis for the SDA Design Standard (2019) figures channel as evaluated in Chapter 5. Deontic force (the classification of a normative statement as obligatory, permissive, or prohibitive) is a foundational property for any compliance verification system. A requirement whose deontic force cannot be determined is effectively unenforceable. The analysis covers the full 406-entry serialised figures corpus and reports deontic distribution by field type, by SDA design category, and across the extracted predicate vocabulary.

Corpus-Level Deontic Distribution

Of the 406 entries in the serialised figures corpus, 205 (50.5%) carry modal content and 201 (49.5%) are non-modal. The non-modal entries comprise context descriptions (67 entries), applicability markers (107 entries), and non-modal notes (27 entries), all of which serve framing, scoping, or informational functions rather than prescriptive ones. The near-even split between modal and non-modal content reflects the structural composition of the serialised corpus, which captures design requirements alongside their surrounding contextual metadata.

Table A5-DF.1: Corpus-level deontic force distribution

|:—————|——:|———–:|

Source: Figures-Channel Deontic Classification, corpus-level force distribution (see Data availability above).

Within the modal entries, the overwhelming majority are obligatory: 189 of 205 modal entries (92.2%) carry obligatory force. Permissive entries (13, or 6.3%) and prohibitive entries (3, or 1.5%) together account for less than 8% of modal content. This distribution confirms that the SDA Design Standard’s figure channel is predominantly prescriptive rather than permissive. In summary, the corpus-level deontic profile establishes a near-unipolar obligation structure that has direct consequences for the implementation logic of automated verification systems. The next section examines the specific modal terms through which this obligatory force is expressed.

Modal Term Inventory

The 205 modal entries employ six distinct modal markers. The following table shows the full inventory with occurrence counts and force class assignments.

Table A5-DF.2: Modal term inventory

|:—————-|————:|:————-|

Source: Figures-Channel Deontic Classification, modal-term inventory (see Data availability above). Total exceeds 205 because some entries contain multiple modal terms.

The marker “shall” dominates with 200 occurrences, accounting for 92.6% of all modal term instances. This monosyllabic indicator is the cornerstone of Australian standards drafting practice, where “shall” signals mandatory compliance per the conventions of Standards Australia and ISO directive usage. Its near-exclusive use in the SDA figure channel simplifies automated deontic classification: a pattern-matching approach that identifies “shall” captures the vast majority of obligatory content. Zero entries were flagged as deontic-ambiguous, confirming that the standard’s modal vocabulary is sufficiently disciplined to support deterministic force classification. Building on this modal term inventory, the next section disaggregates deontic force by field type to reveal the structurally significant finding that all design requirement entries are 100% modal.

Deontic Force by Field Type

The cross-tabulation of deontic force against field type reveals a structurally significant finding: design requirements are 100% modal, while context and applicability entries are 100% non-modal by construction.

Table A5-DF.3: Deontic force by field type

|:——————–|——:|——:|———–:|———–:|———–:|————:|———-:|

Source: Figures-Channel Deontic Classification, force by field type (see Data availability above).

Every one of the 189 design requirement entries carries explicit deontic force: there are no design requirements with ambiguous or absent modal marking. Within the Design requirement field type, 181 entries (95.8%) are obligatory, 7 (3.7%) are permissive, and 1 (0.5%) is prohibitive. The single prohibitive design requirement is ID 235, Figure 10: “Use of shower screens is not permitted.”

Notes present an intermediate profile: 16 of 43 (37.2%) carry modal content. The 8 obligatory notes represent requirements that a text-only parser might overlook if it treats “Note” fields as non-normative. These 8 entries embed prescriptive content in a conventionally informational field type, representing a subtle standards drafting pattern that the serialisation schema correctly captures. Overall, the field-type disaggregation confirms that deontic force is not uniformly distributed and that field-type-aware processing is necessary for complete normative capture. The next section presents the category-level analysis that reveals how this obligatory force concentrates differentially across the four SDA design categories.

Deontic Force by SDA Category

Because a single entry may apply to multiple categories, the disaggregation by SDA design category does not produce exclusive counts. The following table reports obligatory and permissive counts per category.

Table A5-DF.4: Obligatory and permissive entries by SDA category

|:———————–|———–:|———–:|

Source: Figures-Channel Deontic Classification, force by category (see Data availability above). Prohibitive entries are not disaggregated by category in the source data.

The category-level distribution reveals a substantial asymmetry: the Fully Accessible (135 obligatory) and High Physical Support (138 obligatory) categories carry more than three times the obligatory requirement count of Improved Liveability (41) and Robust (38). This differential reflects the SDA framework’s design intent: higher-support categories impose more prescriptive spatial requirements. The near-parity between Fully Accessible (135) and High Physical Support (138) obligatory counts, and between Improved Liveability (41) and Robust (38), further confirms the two-tier pattern observed in Appendix: Polysemy Metrics and Confidence: the four SDA categories effectively operate as two paired tiers at the level of deontic structure. Taken together, the corpus-level, field-type, and category-level disaggregations present a consistent picture of how deontic force is distributed through the standard, and thus establish the informational basis that the predicate vocabulary analysis in the next section extends to the level of specific prescriptive verbs.

Robust-category clause provenance

The four SDA design categories reported here (Improved Liveability, Robust, Fully Accessible, and High Physical Support) are those of the NDIA Specialist Disability Accommodation Design Standard, Version 1 (October 2019). The Robust-category requirements counted above (38 obligatory, 1 permissive) are referenced against that Standard at the category level. The per-clause Robust identifiers carried into the Chapter 10 decisive-instance analysis (the D2-D14 series) are registered here as direct category-level references to the Standard, Version 1, rather than as anchors to numbered per-clause edges: a granular per-clause Robust typology has not been extracted, so per-clause anchoring is left as an open provenance task for the Chapter 10 Robust baseline library rather than asserted here.

Predicate Vocabulary

The 189 design requirement triples employ 17 unique predicate phrases, indicating a controlled but not minimal predicate vocabulary. The following table presents the distribution of the 10 most frequent predicates.

Table A5-DF.5: Predicate vocabulary distribution (top 10 of 17)

|:———————–|————:|——–:|

Source: Figures-Channel Triple Statistics, predicate-vocabulary distribution (see Data availability above).

The top five predicates account for 164 of 189 triples (86.8%), confirming that the SDA figure channel uses a narrow set of prescriptive verbs. The dominance of “shall be” (67 occurrences, 35.4%) and “shall be provided” (41 occurrences, 21.7%) reflects the standard’s focus on spatial provision requirements. The locative predicate “shall be located” (23, 12.2%) captures positioning requirements, while “shall have” (17, 9.0%) captures property assignments. The controlled predicate vocabulary has a direct practical consequence: the 17-predicate inventory is small enough that a compliance verification system can define an explicit handler for each predicate type rather than relying on generalised semantic parsing. Overall, the predicate vocabulary confirms that the SDA figure channel encodes a structurally narrow but semantically dense set of spatial obligations that the serialisation schema can represent with high completeness.

The same concentration appears at the level of relations, not only predicates. Across the SDA corpus the has_quality relation is the most heavily loaded relation, with a Gini coefficient of 0.6639 over its four semantic subgroups (dimensional, performance, presence, and categorical). The has_quality relation is introduced in Chapter 5 (§5.5) and this loading figure is quantified in Chapter 6 (§6.2); it is also catalogued in Appendix F: Data Dictionaries. It is printed here so the corpus’s relation-level load concentration sits alongside the predicate-level distribution above.

Implications for Compliance Verification

The 100% modal rate for design requirements means that every spatial requirement extracted from the SDA figures can be assigned a definitive deontic classification, eliminating one of the most common sources of standards interpretation disputes. The concentration of obligatory force (“shall” accounting for 92.6% of all modal markers) further simplifies the compliance logic. A verification system can adopt a default-obligatory posture and flag only the small minority of permissive (13) and prohibitive (3) entries for differential treatment.

The zero deontic ambiguity count confirms that no manual adjudication is required for force classification in this corpus, supporting the feasibility of fully automated deontic tagging in the serialisation pipeline. Combined with the 100% identity resolution rate reported in Appendix: Chapter 5 Figures-Channel Ambiguity Analysis, these results establish that both the what (entity identity) and the how strongly (deontic force) of every design requirement can be determined without human intervention. Therefore, the deontic and predicate evidence presented in this appendix supports the Chapter 5 evaluation claim that the serialisation schema provides a complete and deterministic normative representation of the SDA Design Standard’s figures channel.

Chapter 5 Polysemy Burden Assessment

The polysemy burden identified in the SDA Design Standard (2019) serialised figures corpus is documented here as part of the evaluation in Chapter 5. Polysemy, the capacity of a single term to carry multiple distinct meanings, constitutes a structural hazard for automated standards interpretation and compliance verification. Three distinct forms of polysemy are reported. WordNet lexical polysemy concerns general-language sense counts. Dimensional polysemy occurs where the same concept carries different quantitative values across SDA categories. Categorical absence polysemy occurs where entities are specified in some categories but absent from others. The analysis draws on 56 unique canonical entities extracted from 189 figure-based design requirement triples.

WordNet Lexical Polysemy

Lexical polysemy is measured at the granularity of the canonical term inventory underlying the corpus analysis (1,236 terms). Each of the 56 canonical entities is resolved to the syntactic head of its concept phrase; degree markers (“minimum”, “maximum”), quantity qualifiers (“of 2200 mm”), and trailing prepositional attachments are stripped before lookup, so that a left modifier or a qualifier token can never supply the sense count for a compound. Under this head-anchored measurement, 26 of the 56 entities have their full name or cleaned head term present in the inventory: 13 are polysemous and 13 are monosemous. The remaining 30 entities carry head terms that fall below the inventory’s frequency threshold; these are reported as not measured rather than imputed, consistent with the dual-reporting norm adopted throughout Chapter 5. The polysemy burden among the measured entities is not uniformly distributed; a small number of entities carry disproportionately high sense counts.

Table A5-PB.1: Polysemous entities in the serialised vocabulary (all 13, head-anchored measurement)

Source: Entity Registry v2 (head-anchored polysemy records), sorted by WordNet sense count descending; foundational primitive per the term-to-foundational-mapping deposited with the data package (see Data availability above).

The entity Tap is the clearest illustration of the polysemy problem. WordNet records 20 senses spanning the physical fixture (“water_faucet.n.01”), the action of striking (“rap.n.02”), wire-tapping (“wiretap.n.01”), and tap dancing (“tapdance.v.01”). In the SDA context, only the plumbing fixture sense is relevant, but an automated system without domain-specific disambiguation would need to evaluate all 20 candidates. The compound Tap_And_Water_Source illustrates the corrected method’s behaviour: its head term resolves to “water source”, which the inventory records as monosemous, so the compound no longer inherits the 20-sense burden of its left conjunct; that inheritance was an artefact of the defective fallback.

The entity Door presents a different polysemy profile: only 5 WordNet senses, but 28 occurrences across the corpus with 20 distinct surface variants. Referential scope is the primary burden for Door: it variously denotes the physical barrier, the doorway opening, the circulation space, and the attached hardware. The serialisation schema resolves all 20 surface forms to a single canonical entity. Underlying referential polysemy remains a challenge for finer-grained semantic analysis. Overall, the head-anchored analysis confirms that domain-specific sense disambiguation is required for the 13 measured-polysemous entities, half of the measurable vocabulary, while the 30 below-threshold entities require inventory extension before any sense-level claim can be made about them. Dimensional polysemy is examined in the next section: the same entity there requires different quantitative values depending on the applicable SDA design category.

Dimensional Polysemy

Dimensional polysemy occurs when the same conceptual entity requires different quantitative values depending on the SDA design category. Five cases of dimensional polysemy are identified in the figure-based corpus.

Table A5-PB.2: Dimensional polysemy cases by entity and SDA category

Source: Figures-Channel Polysemy Analysis, dimensional-polysemy records (see Data availability above). IL = Improved Liveability; Robust; FA = Fully Accessible; HPS = High Physical Support.

These 5 cases are structurally significant because they represent requirements where the same entity in the same spatial context carries materially different dimensional specifications. Graduated dimensional polysemy is illustrated by the entity Door. Improved Liveability and Robust categories require 820 mm minimum clear opening; Fully Accessible requires 900 mm; High Physical Support requires 950 mm. The 130 mm range is substantive: it determines whether a wheelchair passes with one-sided or two-sided clearance. A compliance system must resolve the applicable category before evaluating the dimensional requirement. Failure to do so produces systematic over-specification or under-specification errors.

A structural asymmetry is also confirmed by the dimensional polysemy cases. Improved Liveability and Robust share identical dimensional values; so do Fully Accessible and High Physical Support. This pairing pattern is confirmed independently by the deontic force distribution in Appendix: Chapter 5 Predicate Coverage and Deontic Force. It suggests the four-category SDA classification effectively operates as a two-tier system at the quantitative level. In summary, the five cases require that any compliance system resolve category context before evaluating dimensional conformance. Categorical absence polysemy is examined next: entities there appear in some categories but are entirely absent from others.

Categorical Absence Polysemy

Categorical absence polysemy occurs when an entity is specified in some SDA categories but absent from others. The analysis identifies 36 such cases. Thirty-four of 36 absence cases follow a single pattern: entities appear in Fully Accessible and High Physical Support but are absent from Improved Liveability and Robust.

Three bathroom entities are absent from lower-tier categories: WC_Pan, Hand_Wash_Basin, and Shower. Spatial clearance entities include Knee_And_Toe_Clearance_Zone and Encroachment-Free_Zone. Kitchen specifications (Cooktop, Kitchen_Bench, Pantry, and Drawer-Style_Dishwasher) are absent from lower-tier categories. Bedroom specifications (Bed, Bedroom) are likewise absent from Improved Liveability and Robust requirements. These absences reflect the SDA framework’s graduated approach to physical support prescription.

Washing_Machine displays the inverse pattern: present in Improved Liveability and Robust, absent from Fully Accessible and High Physical Support. Higher-support categories prescribe alternative laundry configurations in its place. Minimum_900_Mm_Wide_Continuous_Accessible_Path is exclusive to High Physical Support, absent from all three other categories.

Conditional polysemy is recorded in one additional case. WC_Pan in the context of “DIM +/- DIM from the rear wall” carries identical values (800 mm +/- 10 mm) for both FA/HPS and an unspecified category. Polysemy is present but produces no divergent requirements in this degenerate case. Overall, the 36 absence cases define each SDA category’s regulatory reach. Automated systems must represent absence explicitly rather than treating it as a missing value.

Aggregate Polysemy Burden and Implications

Forty-two polysemy cases are identified across all three forms: 5 dimensional, 1 conditional, and 36 categorical absence. The affected corpus contains 189 design requirement triples. A polysemy incidence rate of 22.2% results: approximately one in five requirements is subject to context-dependent semantic variation.

Domain-specific sense disambiguation is required for the 13 lexically polysemous entities identified by the head-anchored measurement before WordNet-based or embedding-based analysis can be applied reliably; a further 30 entities are unmeasured at term granularity and must not be treated as either burdened or clear. The serialisation schema addresses this through canonical entity resolution, but downstream systems must recognise that general-purpose NLP tools will systematically over-count semantic complexity.

Category context resolution is required by the 5 dimensional polysemy cases before any automated compliance check evaluates dimensional conformance. The 36 categorical absence cases define each category’s regulatory boundaries. Automated systems must distinguish “this entity is not required” from “this entity was not found.” The serialisation schema records which categories each entity’s requirements apply to, making the distinction explicit. Taken together, the three polysemy forms confirm that polysemy is a structural feature of the SDA standard. The schema must represent it explicitly. This polysemy burden evidence supports the artefact design decisions in Chapter 5.

Chapter 5 Figures-Channel Ambiguity Analysis

This appendix presents the five-dimensional ambiguity analysis of the SDA Design Standard (2019) figures channel as implemented by the serialisation schema evaluated in Chapter 5. The analysis covers the full 406-entry serialised figures corpus and reports the proportion of ambiguity dimensions resolved by the schema relative to those present in the source material, together with interpretation of residual ambiguity in each dimension. The complementary text-channel evaluation is documented in Appendix: Chapter 5 Evaluation Results and Data Package.

Five-Dimensional Ambiguity Model

The resolution rate is computed across five orthogonal dimensions, each capturing a distinct facet of interpretive uncertainty in normative standards text. Each dimension is scored as a binary per entry: present (the dimension is relevant to the entry), resolved (the schema eliminates the ambiguity), or residual (the ambiguity persists despite serialisation).

1. Identity ambiguity: whether the subject entity of a requirement can be unambiguously identified.

2. Structural ambiguity: whether the syntactic decomposition of a requirement into subject-predicate-object triples is deterministic.

3. Deontic ambiguity: whether the modal force (obligatory, permissive, prohibitive) of a requirement is unambiguously classifiable.

4. Dimensional ambiguity: whether quantitative values attached to a requirement are fully specified.

5. Categorical ambiguity: whether the SDA design categories to which a requirement applies are explicitly determined.

Per-Dimension Results

The overall resolution rate across the figures corpus is 0.9148, indicating that 91.48% of all ambiguity-bearing dimensions are resolved through the serialisation process. The per-dimension results are presented in the following table.

Table A5-FA.1: Figures-channel resolution rate by dimension

|:————-|——–:|———:|———:|—————-:|

Source: Figures-Channel Resolution-Rate Scores, global dimension summary (see Data availability above).

Two dimensions achieve full resolution; one achieves near-perfect resolution; two dimensions carry characterised residual ambiguity. The pattern is structurally predictable and reflects properties of the source standard rather than limitations of the schema design. In summary, the five-dimensional profile establishes a clear resolution hierarchy that the next sections document in detail, proceeding from perfect-resolution dimensions through near-perfect resolution to the two dimensions where characterised residual ambiguity remains.

Perfect-Resolution Dimensions

Identity ambiguity is resolved at 100.00% across 174 instances. This outcome reflects the entity-resolution pipeline’s capacity to canonicalise variant surface forms into a controlled vocabulary of 56 unique canonical entities. The 100% resolution rate confirms that every entity referenced in the design requirement entries of the serialised corpus can be unambiguously traced to a single canonical identifier. This is a non-trivial outcome: the source standard deploys 20 distinct surface variants for the entity Door alone, ranging from “Door circulation spaces” to “required internal door circulation space”, yet the schema resolves all of them to the canonical subject without information loss.

Deontic ambiguity is likewise resolved at 100.00% across all 205 modal entries. The modal term inventory recovered from the corpus comprises six distinct markers: “shall” (200 occurrences), “is permitted” (7), “may” (4), “are permitted” (2), “shall not” (2), and “not permitted” (1). Zero entries were flagged as deontic-ambiguous, meaning every requirement with modal content could be assigned to exactly one of the three force classes (obligatory, permissive, prohibitive) without adjudication. The absence of deontic ambiguity is partly attributable to the SDA standard’s disciplined use of “shall” as its obligatory marker: a convention that, while not universally observed in Australian standards drafting, is sufficiently consistent in this instrument to permit automated classification. This implies that automated deontic tagging is fully feasible for this corpus without any manual adjudication phase. The next section documents the single residual in the structural dimension.

Near-Perfect Resolution: Structural

Structural ambiguity is resolved at 99.47%, with 188 of 189 design requirements successfully decomposed into subject-predicate-object triples. The decomposition quality audit reports 187 clean decompositions (98.9%), 1 fallback decomposition (0.5%), and 1 failed decomposition (0.5%). The single failed entry is ID 235, sourced from Figure 10: “Use of shower screens is not permitted.” This entry resists triple decomposition because its syntactic structure is a bare prohibition without a clear spatial predicate: the prohibition applies to an entire class of objects rather than to a spatial relationship. The schema records this as a tracked residual rather than discarding the entry, preserving the normative content while acknowledging the structural limitation.

Residual Ambiguity: Dimensional

Dimensional ambiguity is resolved at 88.74%, with 17 residuals across 151 dimensionally-active entries. The residual rate of 12.26% arises from two distinct mechanisms. First, some entries specify dimensional ranges rather than point values (for example, “900 mm to 950 mm based on design category”), where the precise applicable dimension depends on the SDA category context. The serialisation schema captures the range but cannot collapse it to a single value without category resolution. Second, entries referencing external standards (for example, “as per AS1428.1”) inherit dimensional requirements that are not reproduced within the SDA standard’s own text or figures. These external references constitute genuine residual ambiguity from the perspective of the SDA corpus alone: a practitioner must consult the referenced instrument to determine the applicable dimension.

The 17 dimensional residuals are concentrated in the Design requirement field type, where the class profile reports an average ambiguity delta of 0.949, still high, but acknowledging that approximately one in six dimensionally-active requirements carries some unresolved quantitative specification. This finding has direct implications for automated compliance checking, where dimensional completeness is a prerequisite for programmatic verification. Building on this characterisation of dimensional residuals, the next section addresses categorical ambiguity, which presents the largest residual proportion of the five dimensions.

Residual Ambiguity: Categorical

Categorical ambiguity presents the lowest resolution rate at 72.84%, with 63 residuals out of 232 categorically-active entries. The 63 residuals divide into two structural sources with distinct interpretations.

The first source accounts for 43 residuals (68.3% of categorical residuals) and originates in the Note field type. Notes in the serialised corpus carry categorical tagging (all 43 Note entries are categorically-active) but none are categorically-resolved, producing a categorical resolution rate of 0.0% within Notes. This is by design: notes are informational glosses that contextualise requirements without themselves being category-specific obligations. The schema correctly identifies categorical presence (notes apply within a categorical context) while marking resolution as impossible (notes do not prescribe for a specific category).

The second source accounts for 20 residuals and originates in the Design requirement field type, where 189 categorically-active entries resolve to 169 with 20 residuals. These 20 residuals arise from requirements that span multiple categories with category-dependent dimensional variation: the polysemy cases documented in Appendix: Polysemy Metrics and Confidence. For instance, the entity Door carries a minimum clear opening of 820 mm for Improved Liveability and Robust categories, 900 mm for Fully Accessible, and 950 mm for High Physical Support. The schema serialises each variant but cannot collapse the requirement to a single category without losing the category-contingent dimensional specification. Taken together, the two categorical residual sources account for all 63 unresolved cases in a fully characterised manner, with no unclassified residuals.

Implications for Standards Interpretation

The 91.48% overall resolution rate demonstrates that a structured serialisation schema can eliminate the vast majority of interpretive ambiguity in a building standards instrument. The residual ambiguity is not randomly distributed but is concentrated in two predictable and well-characterised zones: categorical indeterminacy in notes (a genuine semantic property of informational text) and dimensional-categorical coupling (where the same spatial requirement takes different quantitative values depending on the SDA design category). Both forms of residual ambiguity are tracked rather than hidden: the schema flags them explicitly, enabling downstream consumers to route these cases to human adjudication or category-specific lookup tables.

Identity, structural, and deontic ambiguity are resolved at or near 100%. These results establish that the serialisation schema provides a reliable foundation for automated triple extraction. Category-contingent dimensional specification remains inherently complex, but the schema acknowledges this explicitly. Overall, the 91.48% aggregate resolution rate confirms that the schema meets the performance threshold for Chapter 5 evaluation claims. Characterised residuals provide the transparency needed for downstream consumers to route unresolved cases to appropriate adjudication mechanisms.

Chapter 5 Cross-Channel Validation Results

The cross-channel validation analysis for Chapter 5 establishes the complementarity and non-redundancy of the text and figures channels of the SDA Design Standard (2019). The analysis matches figure-derived design requirement triples against text-derived design requirements to classify each figure triple by the degree to which it is confirmed, corroborated, or unmatched in the text channel. The core evaluation claim, that the two-channel architecture is warranted by measurable complementarity rather than by convenience, is grounded in these results.

Match Type Distribution

The bidirectional cross-validation produces a three-way match-type classification by comparing each of the 189 figure-based design requirement triples against the 140 text-based design requirements extracted from the standard’s clause structure. Matching employed both concept matching (entity-name overlap) and dimension matching (shared quantitative values) to classify each figure triple into one of three match types.

Table A5-CC.1: Cross-validation match type distribution

Source: Text-Figure Cross-Validation Results, match-type summary (see Data availability above).

The most consequential finding is that 73 figure-based triples (38.6%) are classified as figure_only: they carry design requirements that exist exclusively in the figure channel with no recoverable textual counterpart. These are not layout details or graphical annotations; they are dimensioned, deontic requirements with subject-predicate-object structure and obligatory modal force. A practitioner or compliance system relying solely on the text clauses of the SDA standard would miss 38.6% of the figure-originated normative content. In summary, the figure-only classification rate is not a marginal residual but a primary structural characteristic of the standard that must inform every downstream compliance system design decision.

The 77 concept-matched triples (40.7%) are defined as figure triples that share entity-level vocabulary with text requirements but lack dimensional confirmation. The text channel acknowledges the existence of the entity and some associated requirement, but the specific dimensional values prescribed in the figure are not reproduced in the text. For example, a text clause may reference “gate clearances” in general terms while the figure specifies the exact 820 mm minimum opening width. The concept match confirms thematic alignment without dimensional parity.

The 39 dimension-matched triples (20.6%) provide the strongest form of cross-channel confirmation: both the entity reference and at least one dimensional value appear in both channels. These cases provide the highest confidence that the figure and text channels encode the same normative intent, albeit often with different levels of specificity.

Directional Asymmetry

The cross-validation establishes a pronounced directional asymmetry between the two channels. Of the 140 text-based design requirements, only 19 (13.6%) have a confirmed figure-based match. The remaining 121 text-based requirements (86.4%) exist only in the text channel. This asymmetry is not symmetrical with the figure-only finding: while 38.6% of figure triples lack textual counterparts, 86.4% of text requirements lack figure counterparts.

This directional pattern is structurally explicable. The text channel of the SDA standard covers a broader regulatory scope (including procedural requirements, cross-references to external standards, and conditional applicability rules) that are not amenable to diagrammatic representation. The figure channel, by contrast, concentrates on spatial and dimensional requirements that benefit from graphical expression. The result is that the two channels are complementary rather than redundant: the text channel carries procedural and conditional content, while the figure channel carries spatial-dimensional content, with a narrow overlap zone of approximately 19 shared requirements. Overall, the directional asymmetry confirms that neither channel is derivable from the other, and that any compliance system must treat both as primary rather than treating either as a derivative supplement. The next section examines the design implications for schema architecture that follow from this structural finding.

Implications for Schema Architecture

The cross-validation results establish that neither channel is a subset of the other. A practitioner relying exclusively on the text channel would miss 38.6% of the figure-originated normative content. A practitioner relying exclusively on the figure channel would miss 86.4% of the text-originated normative content. Only a two-channel extraction architecture that processes both sources and integrates their outputs into a unified triple store can claim to represent the full normative scope of the SDA Design Standard. This cross-channel reconciliation is the integration measure registered as XCC-INT-03: the measure, governed by the pre-registered cross-channel resolution threshold of Section 4.5, of how completely the two-channel architecture recovers the normative content that either channel read alone would omit.

These results establish the warrant for the design decision to develop parallel serialisation pipelines for the text and figures channels rather than extracting from one channel alone. The figures-only requirements (73 triples) include specifications for fundamental accessibility features (door opening widths, ramp landing dimensions, bathroom fixture clearances) whose omission from a compliance system would constitute a material failure of the regulatory representation.

The 19 dimension-matched triples also provide an important verification function. Their presence confirms that the two extraction pipelines, operating independently on different source modalities (text versus figure diagrams), converge on the same normative content for a subset of requirements. This convergence provides an internal consistency check on the extraction methodology that does not require external ground truth. Taken together, these three match-type findings (38.6% figure-only, 40.7% concept-matched, and 20.6% dimension-matched) establish that the two channels are genuinely complementary rather than partially overlapping duplicates, and therefore that the two-channel architecture is the minimal design sufficient to capture the full normative scope of the standard. This establishes the evidentiary basis for the schema design decisions in Chapter 5 and the reproducibility verification procedures documented in the companion data package appendix.

Chapter 5 Data Pipeline and Reproducibility

The data pipeline architecture, versioned file inventory, and verification procedures supporting the quantitative findings in Chapter 5 are documented here. Reproducibility is a foundational requirement for design science research artefacts: the serialisation schema and its associated analysis pipeline must be transparent in their inputs, transformations, and outputs such that an independent researcher could, given access to the same source materials, reproduce the findings reported in the evaluation. This appendix also identifies the methodological limitations that constrain full reproducibility. The full data package, including all analysis scripts and versioned outputs, is deposited as the Chapter 5 artefact bundle described in Appendix: Chapter 5 Evaluation Results and Data Package, where the deposit location and canonical artefacts are enumerated.

Pipeline Architecture

The analysis pipeline proceeds through four stages, each producing versioned outputs that serve as inputs to the subsequent stage.

Stage 1: Source Extraction. The SDA Design Standard (2019) produces two parallel extraction channels: a text channel extracting clause-based requirements from the standard’s prose (611 total entries, 140 classified as design requirements), and a figure channel extracting diagram-based requirements from the standard’s 19 figures (48 sub-figures; 406 total entries, 189 classified as design requirements). Both channels preserve the full source text alongside structured metadata (figure references, clause references, field types, applicability markers).

Stage 2: Normalisation and Triple Extraction. Figure-channel entries are encoded in a structured, consistently field-typed and applicability-tagged form. Design requirements were decomposed into subject-predicate-object triples using pattern-based extraction, producing 189 triples with 56 unique canonical entities and 17 unique predicate phrases. Entity resolution was performed to map variant surface forms to canonical identifiers; for example, 20 surface variants of “Door” were mapped to the canonical entity Door. The decomposition quality audit reports 187 clean extractions (98.9%), 1 fallback (0.5%), and 1 failure (0.5%).

Stage 3: Multi-Dimensional Analysis. Four parallel analyses are applied to the normalised corpus: (a) five-dimensional resolution-rate scoring (documented in Appendix: Chapter 5 Figures-Channel Ambiguity Analysis); (b) deontic force classification and inventory (documented in Appendix: Chapter 5 Predicate Coverage and Deontic Force); (c) polysemy assessment (documented in Appendix: Chapter 5 Polysemy Burden Assessment); and (d) text-figure cross-validation (documented in Appendix: Chapter 5 Cross-Channel Validation Results). Each analysis was executed by a dedicated, version-controlled analysis program, producing both a machine-readable data export and a human-readable narrative report.

Stage 4: Unified Data Assembly. The outputs of Stage 3 are assembled into a single unified explorer dataset containing all 56 entities, 189 triples, and associated metadata in one queryable structure. This dataset serves as the canonical data source across all four evidence domains. An interactive, self-contained browser explorer was generated from it to support entity browsing, triple inspection, and design-category filtering. In summary, the four-stage pipeline produces a transparent, versioned chain of evidence from source standard to quantitative findings, and the next section describes the named datasets that make each stage auditable.

Named Datasets by Pipeline Stage

The tables below describe the named source, intermediate, and output datasets produced at each pipeline stage, each with its record scope and approximate data volume. All datasets are deposited in the Chapter 5 artefact bundle whose deposit location is recorded in Appendix: Chapter 5 Evaluation Results and Data Package; the bundle’s own data dictionary documents the field-level schema of each.

Stage 1: Source Datasets

Stage 2: Normalised Datasets

Stage 3: Analysis Datasets

Each Stage 3 analysis was produced by a dedicated, version-controlled analysis program documented in the bundle’s data dictionary.

Stage 4: Unified Dataset

Verification Procedures

Each pipeline stage incorporates internal verification that allows the findings to be cross-checked against multiple independent representations of the same underlying data.

Stage 1 Verification. The figure extraction was validated against source images. A figure-extraction comparison record, deposited in the bundle, documents discrepancies between the initial and revised extractions, with all discrepancies resolved in the revised pass.

Stage 2 Verification. The normalisation stage produces a statistics dataset that independently reports field-type distributions (context: 67, Design requirement: 189, Applicable to: 107, Note: 43, total: 406), enabling cross-check against the raw extraction counts. A normative-voice audit within this dataset confirms that 181 design requirements contain “shall” and 8 contain “permitted,” with zero design requirements lacking a normative verb.

Stage 3 Verification. Each analysis produces both a machine-readable data export and a human-readable narrative report. The narrative reports present the same quantitative findings in prose, enabling manual review against the data exports. The parity assessment explicitly tracks gaps between the text and figure analysis channels, recording three gaps closed and three remaining minor gaps. Building on these stage-level checks, Stage 4 verification confirms that the unified dataset preserves a complete provenance chain from source material to final output.

Stage 4 Verification. The unified dataset is built by a deterministic assembly procedure that reads from the Stage 3 datasets. The dataset’s metadata records source identification, entity count (56), triple count (189), and version, providing a provenance chain from source to final output. Every quantitative claim in the evaluation appendices is traceable to a named field in a named dataset described above.

Limitations on Reproducibility

Five limitations constrain the reproducibility of this analysis. An independent researcher attempting replication should be aware of each.

Manual entity resolution. The 56-entity vocabulary records a specific set of resolution decisions that are defensible but not uniquely determined, because the mapping of 20 surface variants of “Door” (and similar variant sets for other entities) to canonical identifiers requires manual adjudication. A different researcher might draw canonical boundaries differently, for instance, treating “Door circulation spaces” and “Door sizes” as distinct entities rather than variants of Door. The 56-entity vocabulary reflects a specific set of resolution decisions that are defensible but not uniquely determined.

WordNet polysemy scope. WordNet sense inventories provide general-language sense counts but do not capture domain-specific technical meanings. An entity like Bollard is flagged as monosemous despite having a specific access-control meaning in the SDA context, while entities like Tap inherit 20 general-language senses mostly irrelevant to building standards. Polysemy burden figures should be interpreted as upper bounds on general-language ambiguity rather than as precise measures of domain-specific ambiguity.

Source standard access. The SDA Design Standard (2019) is a copyrighted instrument administered by the National Disability Insurance Agency. The source document cannot be redistributed with the research materials. An independent researcher would need to obtain it through authorised channels to reproduce the Stage 1 extraction.

Extraction subjectivity. The figure extraction process requires interpreting graphical annotations, dimensional callouts, and spatial relationships from rasterised diagram images. While the extraction prompt template and format specification provide detailed instructions, some degree of interpretive judgement is inherent in reading dimensions from scaled diagrams.

Pipeline iteration history. The text channel produces 7 analysis iterations (cycles 0-6), while the figure channel produces a base analysis plus a deepening pass. All final outputs derive from a single consistent extraction snapshot, so that the deposited datasets form one coherent version for verification. Earlier intermediate outputs may differ from the final versions. Overall, these five limitations bound the reproducibility claim in a transparent manner: they identify where independent replication requires either access to the source standard, tolerance of alternative resolution judgements, or awareness of WordNet’s general-language scope. The next section summarises the pipeline’s aggregate scope and the traceability structure that holds across all four evidence domains.

Summary

The data pipeline from source standard to evidence appendices is traceable across 4 stages and roughly two dozen versioned datasets totalling approximately 6.9 MB of structured data. Every quantitative claim in the evaluation appendices is traceable to a named field in a named dataset, and every analysis is executed by a documented, version-controlled program operating on versioned inputs. Within the constraints identified above, the pipeline provides a transparent and auditable chain of evidence from source material to quantitative findings. Therefore, the reproducibility infrastructure documented here meets the transparency requirements that design science research imposes on artefact-based evaluation and supports the Chapter 5 claim that the evaluation findings are independently verifiable within the stated constraints.