The method
Transparent end to end, from the authority gate to the delta re-run. Every finding carries its evidence, and every score discloses its method.
Nothing is extracted without documented client authority. The tiers ladder by data sensitivity: Tier 1 touches configuration only and needs the lightest gate; Tiers 3 and 4 reach matter and party data and need full authority. Production legal data is never extracted before that authority is in place.
Each corpus (work items, key dates, attributes, parties, connections, and notes) is pulled through the API using stable (date, identifier) boundaries, deduplicated, and reconciled against platform totals before any headline figure is quoted.
Embeddings and language models run on our own on-premise hardware in Australia: a local model on a machine we own and operate. Nothing is sent to a third-party AI service. Where a review runs in-boundary instead, that same stack is deployed inside your environment and the analysis does not cross a border at all.
Every finding names its extraction source and as-of date. Observed measures (counts and dates) are kept separate from model-derived inferences (entity matches, cluster labels, extracted reserves), and each inference carries its validation status.
Each domain emits a score out of 100 from transparent weighted indicators with published green/amber/red thresholds. The full methodology (values, thresholds, and weights) is disclosed on every report. The scoring method is ours; the evidence is yours.
Under the Pulse subscription the scripted pipeline re-runs on a schedule. You buy the delta (what changed since last quarter) and a health-score trend the board can watch. Every re-run defends the last quarter's remediation.
Honesty markers
Counts and dates are observed measures. Entity matches, cluster labels, and extracted reserves are model-derived inferences, presented with their validation status, never dressed up as fact.
Acted-both-sides hits are screening candidates for your authoritative conflicts process. We surface them for a human to decide; we never decide them.
Each domain score is a weighted average of indicators with published green/amber/red thresholds. The full methodology, with its values, thresholds, and weights, appears on every report.
The methodology is validated end-to-end on one live engagement. We treat generalisations as hypotheses until more engagements confirm them, and we say so.
