Private data cannot drift into a broader audience.
If CRM data is Operator-only, a workflow cannot write it to a public channel, return it to an uncleared caller, or send it to an unapproved model.
Team members can create their own AI workflows inside a signed governance envelope. You define which data may reach which people, models, and destinations—and the checker rejects workflows outside it.
The people closest to the work should be able to automate it. They should not need—and do not receive—the power to rewrite the organization’s trust boundaries.
Authors connect approved resources, test the workflow, read denial diagnostics, restructure unsafe flows, redact records, and file precise escalation requests.
Governance labels real resources, sets provider and human clearances, and decides whether a narrow declassification or endorsement crossing is justified.
Information-flow control is not just access at the front door. Labels derive through joins and survive handoffs, facts, child workflows, model prompts, tool results, and final outputs.
If CRM data is Operator-only, a workflow cannot write it to a public channel, return it to an uncleared caller, or send it to an unapproved model.
Internet or inbox content cannot shape a higher-integrity sink unless it passes through an explicit, authorized endorsement boundary.
An agent serving a user is denied every read above that user’s clearance. Provider clearance is independently checked for all context sent to a model.
The signed envelope names real file stores, channels, providers, humans, and readers. Workflows are checked against that shared policy before they run.
CRM, inbox, SMTP, file stores, model providers, and principals get confidentiality and integrity declarations.
People implement their own workflows and repair denied designs without waiting for central engineering.
Every rule is checked. Unlabelled resources fail closed as flagged risks. The report states exactly what was proven.
Some releases are legitimate. WhippleScript gives them a narrow, source-marked path instead of turning one exception into ambient permission.
redact … keep creates a value with only the named fields. Dropped fields are absent both statically and at runtime; downstream code cannot recover them.
A marked coercion can cross confidentiality only with a matching signed grant. Its output schema bounds what leaves, and the crossing appears in the report.
A marked judgment can cross the integrity boundary under a matching grant. Raw low-integrity data still cannot write trusted state.
Low-integrity input is denied the ability to steer its own declassification. The escape hatch’s control path is governed too.
A governance agent signs the envelope. The signed artifact carries a SHA-256 attestation over policy content; edited content fails verification instead of quietly changing the claim.
The guarantee report lists proven invariants per resource, violations caught, ungoverned risks, cleared principals, and every accepted declassify or endorse crossing in one trusted-surface section.