Check before sending protection.
The authorized connector rechecks authority start and expiry times when dispatching a protection request. An approval that has expired cannot be reused merely because it was valid at session startup.
The original MVP, accepted Capability 1 and two bounded runtime corrections have been brought together and tested. The evidence remains controlled and synthetic; real-company operation has not been established.
Today’s verification matched the accepted October 7 baseline plus exactly three approved runtime file changes. All 18 bounded regression tests passed again on Windows. The configured protection test and fictional management-escalation test also passed through the browser.
The authorized connector rechecks authority start and expiry times when dispatching a protection request. An approval that has expired cannot be reused merely because it was valid at session startup.
Within a running session, evidence accumulates across reads for each work item. Previously known requirements are reevaluated when the evidence stream is quiet, so a deadline can become overdue without a new record arriving.
Requested, acknowledged, applied, consumed and effect observed remained separate. The fictional management test separately recorded notice requested, delivered, received and management acted.
Next step: complete company-facing Capability 1 setup and verify one bounded, authorized company connection. The accepted October 7 checkpoint and original frozen MVP remain unchanged.
A company can define the evidence a consequential workflow is expected to produce, including its source, conditions, timing and work-item identity. Josephine can evaluate whether that expected evidence is present — and preserve uncertainty when it is not.
For company-defined evidence requirements, Josephine can now distinguish several materially different conditions instead of collapsing them into a simple present-or-missing answer.
That allows Josephine to look not only at what did happen, but also at whether required evidence that should exist by now can actually be established.
Josephine does not invent what evidence should exist. The company defines the expected source, conditions, timing and meaning.
Controlled testing demonstrated separate requirement evaluation for multiple work items appearing in the same evidence stream.
Missing anchors, conflicting conditions or unusable timestamps are preserved as unresolved rather than converted into invented certainty.
Fresh regression testing preserved the existing configured protection lifecycle and management-escalation behavior.
Requested, acknowledged, applied, consumed and effect-observed evidence remain separate stages.
Required evidence being absent does not by itself establish whether an AI agent, automation, human action or another condition caused the problem.
The controlled engineering question has advanced. The next meaningful question is whether the same evidence discipline holds against one bounded, company-authorized workflow using real company evidence sources.