Evidence discipline
Know which parts are facts, estimates, interpretations or still unknown.
Coral Shark keeps your description, measured information, estimates and AI-assisted interpretations separate so a plausible answer never becomes false certainty.
The evidence classes you will see
Example: workload
If a customer says a job happens about 14 times per working week, 48 weeks per year, and takes about 41 active minutes each time, those inputs remain customer estimates. Software can then calculate the represented annual occurrences and active hours reproducibly.
The result is not automatically a cash saving.
Example: workflow structure
A customer may describe the process in ordinary prose. AI can propose a structured step sequence from that wording, but those steps remain inference unless separately confirmed or otherwise evidenced.
An inference based on an unknown field is rejected.
What counts as a customer-facing recommendation
A recommended change must be supported by information we can point to. If the basis is unknown, the Brief says what needs checking or gives a supported reason not to act.
Plausible language without a supported recommendation fails the quality check.
Released capacity, estimated value and cash saving are different things.
A Workflow Fix has its own stricter scope and safety check.
Customer proof is a separate permission decision.
A measured outcome can be retained for service learning without becoming marketing proof. Public use requires explicit permission; anonymous permission cannot expose identity, and named permission must record the permitted display name.
Until real permissioned customer outcomes exist, Coral Shark uses transparent methodology and clearly labelled synthetic examples rather than invented testimonials, logos or result claims.