How language models fit industrial energy auditing
A literature synthesis of six gaps between model output and decision-grade audit reports.
My contribution
Synthesized application areas and architectural gaps, emphasizing retrieval, calculations, and human review.
From use cases to missing capabilities
The review maps language-model applications across building energy modeling, fault diagnosis, survey support, recommendation prioritization, standards retrieval, field inspection, and report generation. Six gaps organize the analysis: calculation faithfulness, heterogeneous-evidence retrieval, template adherence, source-grounded citation, field-evidence fusion, and human accuracy governance.
What was reviewed
The reported selection flow starts with 265 records, removes 19 duplicates, screens 246 records, and assesses 105 full texts. Forty-four full texts are excluded, leaving 61 included papers. The application-class map covers 38 papers, a different denominator.
Architectural implication
A useful audit workflow needs traceable source evidence, explicit numeric operands, programmatic checks, and reviewer control. This synthesis motivates AuditGuard but is a separate scholarly record. Its qualitative rubric must not become a numerical performance leaderboard. Keyword and coauthorship networks require approved graph exports before a faithful interactive reconstruction.
System architecture
Select a step, or a block in the drawing, to see its input, computation, output and source.
How the review corpus was selected
A literature synthesis, not a model benchmark. Records narrow from search to full-text review to the included corpus, and a subset forms the application map.
- Selection stage
- Removed
- Included corpus
Identified: 265 recordsInput
- Input
- Literature database searches.
- Computation
- Records identified by the search.
- Output
- 265 records.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Duplicates removed: 19Excluded
- Input
- 265 records.
- Computation
- Duplicate removal.
- Output
- 19 duplicates removed.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Screened: 246 recordsProcessing step
- Input
- 265 minus 19 duplicates.
- Computation
- Title and abstract screening.
- Output
- 105 records advance to full-text review.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Full texts assessed: 105Processing step
- Input
- Screened records.
- Computation
- Full-text eligibility review.
- Output
- 61 included, 44 excluded.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Full-text exclusions: 44Excluded
- Input
- 105 full texts.
- Computation
- Exclusion at full-text review.
- Output
- 44 excluded.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Included: 61 papersOutput
- Input
- 105 minus 44 exclusions.
- Computation
- Synthesis corpus.
- Output
- 61 included papers.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2
Application map: 38 papersOutput
- Input
- Included papers.
- Computation
- Application-class mapping on a subset.
- Output
- 38 papers; a different denominator from the 61 included.
- Source
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 10
Six architectural gaps that organize the synthesis
- Calculation faithfulness
- Heterogeneous-evidence retrieval
- Template adherence
- Source-grounded citation
- Field-evidence fusion
- Human accuracy governance
Adapted from Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2; Figure 10; printed gap-to-pipeline figure caption.
Literature-wide counts describe the reviewed corpus, not personal project impact. Arithmetic is checked at build time.
Animated pulses show conceptual signal flow only. They do not show measured execution time or a live inference run.
Evaluation Scope & Methodology
Samples, denominators, uncertainty, sources and limits
Approach
Preserve literature-selection denominators and distinguish synthesis judgments from measured system performance.
Cohorts and samples
| Cohort | Sample | Definition |
|---|---|---|
| Literature selection flow | 265 identified records | 19 duplicates; 246 screened; 105 full texts; 44 full-text exclusions; 61 included papers. |
| Application-class mapping | 38 papers | Subset used for the application-class map, not the full included corpus. |
Reported measures
| Measure | Unit | Cohort | Denominator | Uncertainty | Evidence |
|---|---|---|---|---|---|
| Literature selection counts | Count | Literature selection flow | Reported selection stages; counts describe sequential pools, not additive categories | None reported | Reported in source |
Limitations
- This is a literature synthesis, not a new model-accuracy benchmark.
- The 38-paper application map is a subset of the 61 included papers.
- The newest supplied version has multiple authors; the older public review has different authorship.
- Qualitative comparison graphics are not empirical accuracy measurements.
Sources
- Large language models in industrial energy auditing: current state, architectural gaps, and the path to programmatic accuracy, Figure 2; Figure 10; printed gap-to-pipeline figure caption (Supplied manuscript reviewed November 2026). Public release pending