Provenance
Data & sources.
Every figure in this report is computed directly from the synthetic demonstration corpus below: a fictional firm's book, generated to exercise every lens with a known, verifiable ground truth.
What was analysed
| Corpus element | Count |
|---|---|
| Matters | 30,000 |
| Instructions | 8,000 |
| Parties | 60,046 |
| Connections | 150,666 |
| Keydates | 110,587 |
The synthetic demonstration corpus, by element.
How it was computed
Method
- Configuration & lifecycle: form-field fill rates, phase ages, and standards alignment computed via pandas over the corpus CSVs (the productised pipeline,
insights.audit). - Matter taxonomy & universe: all-MiniLM-L6-v2 sentence embeddings, UMAP 3D projection, KMeans clustering, and cluster labels from a local LLM (SGLang), all on-premise.
- Entity resolution, limitation triage, workload, conflict radar: direct queries over the corpus, proven on a live client engagement and being generalised into the productised pipeline.
On real engagements: analysis runs on our own on-premise hardware in Australia, nothing is sent to a third-party AI service, and only aggregates and evidence trails appear in deliverables. Where a review runs in-boundary, that stack is deployed inside your environment instead. This report is entirely synthetic data; no client information is represented anywhere in it.
Continue the reportNext: Method & evidence →