Evidence Corpus Profile

The screening matrix clarifies scarcity. Bidirectional capability in the included set was mixed: 15 “yes”, 19 “maybe”, 7 “no”. The dominant exclusion signature was the absence of both bidirectionality and explicit text intermediates.² Overall, the evidence funnel reveals that the field’s availability is broad at the search level but narrow at the methodological level relevant to this thesis’s concerns. The next section characterises the representation families that dominate the included set.

graph TD
  A[Search query scope<br/>Textual floor plan representations] --> B[Screened records<br/>n = 574]
  B --> C{Holistic screening judgement}
  C --> D[Excluded<br/>n = 533]
  C --> E[Included<br/>n = 41]
  E --> F[Extracted evidence fields<br/>method, format, fidelity, interpretability, interoperability, downstream tasks]
  F --> G[Comparative synthesis for Ch2.6.3]
  D --> H[Common exclusion signature:<br/>unidirectional only + no text intermediate representation]

Directionality counts reinforce this mixed maturity. Thirteen records were tagged as bidirectional, 9 as floor-plan to text, and 7 as text to floor-plan. The remainder were either graph-to-layout specialisations or unspecified conversions.³ The two bidirectional figures measure different things and are consistent: the 15 “yes” above records bidirectional capability, whether a method can convert in both directions in principle, while the 13 here records the directionality each study actually demonstrates, so the two-record gap reflects methods capable of, but not shown performing, round-trip conversion. Taken together, the screening and directionality profiles confirm that bidirectional, governance-capable text representations remain a minority practice in the current literature.

Representation Families

Regression governance remains a recurring gap. Few studies publish stable encode-decode regression suites that can be rerun when models, dependencies, or schema bindings change. This matters for lifecycle adaptation because governance workflows are longitudinal and must remain checkable under version drift, not only under one-off experimental conditions.⁵ In summary, none of the three dominant approaches (graph-based, NLP/LLM-based, or token/sequence-based) individually satisfies the full property set. Each approach covers some governance requirements but leaves others unaddressed. Therefore, this evidence profile grounds the Chapter 2 argument that a purpose-designed planimetric notation is needed rather than an extension of an existing family.